User tag fusion method and system based on artificial intelligence

Abstract

The invention relates to the technical field of label fusion, in particular to a user label fusion method and system based on artificial intelligence, and the method comprises the steps: obtaining user labels, carrying out the attention weighted aggregation and image volume accumulation combination of the user labels, and generating a high-order fusion label representation; performing multi-label sorting classification on the high-order fusion label representation, and generating a refined label set; performing user group co-occurrence clustering to generate a context consistent label set; and based on a self-supervised comparative learning model, performing self-supervised comparative learning according to the context consistent label set to obtain an optimized label set. Therefore, the matching degree and relevance between the user and the labels are improved by performing coupling enhancement between the labels and considering the cooperation and conflict relation, the labels can be output according to importance or relevance sorting by introducing a label sorting mechanism, the label quality is improved, the diversity of the labels is ensured, and the user experience is improved. The semantic redundancy of the label set is reduced, and the redundancy is reduced.

Description

Technical Field

[0001] The present application relates to the field of tag fusion technology, and in particular to an artificial intelligence-based user tag fusion method and system. Background Art

[0002] In multi-source, heterogeneous information systems, user tags are essential foundational data for key intelligent tasks such as recommendation systems, user profiling, and precision marketing. Since users generate a large amount of tagged data across multiple platforms and scenarios, it is necessary to fuse these tags to integrate tags from diverse sources, granularity, and expression methods, and to construct a semantically unified and structurally optimized tag set through computational modeling.

[0003] In the prior art, an invention patent with a publication date of May 13, 2025 and publication number CN119992272A discloses an image label generation method based on feature fusion, including: acquiring image data, and generating a multi-scale local feature map through multi-scale feature extraction, extracting category attention information and constructing a category matrix; generating a gating signal through a gating unit, adjusting the feature map, extracting global category features and local category features, and fusing the two to generate a category attention guided feature map; weighting the fused feature map to finally generate a category label map.

[0004] Although the above scheme improves the accuracy and adaptability of image segmentation through multi-scale feature extraction, dynamic adjustment of category attention information, and fusion of global and local features, it still has problems, specifically, the inadequacy of user label sorting and refined classification: the existing multi-label classification method cannot simultaneously consider the semantic similarity, semantic hierarchy and sorting importance between labels, resulting in semantic redundancy in the generated label set and low consistency between users and labels, which cannot meet the requirements of high-quality label fusion. Summary of the Invention

[0005] Based on this, it is necessary to provide an artificial intelligence-based user label fusion method and system that can enhance the coupling between labels through a label scoring model, take into account the collaborative and conflicting relationships, improve the matching and correlation between users and labels, and introduce a label sorting mechanism to output labels in order of importance or relevance, thereby improving the quality of labels while ensuring the diversity of labels, thereby reducing the semantic redundancy of the label set and reducing redundancy.

[0006] The technical solutions of the present invention are as follows: A user tag fusion method based on artificial intelligence, the method comprising: Obtain user tags, perform attention-weighted aggregation and graph convolution aggregation on the user tags, and generate a high-order fused tag representation; Performing multi-label sorting and classification on the high-order fusion label representation and generating a refined label set; Performing co-occurrence clustering of user groups based on the generated refined tag set to generate a context-consistent tag set; Based on a preset self-supervised contrastive learning model, self-supervised contrastive learning is performed according to the context-consistent label set to obtain an optimized label set.

[0007] Specifically, the high-order fusion label representation is subjected to multi-label sorting and classification, and a refined label set is generated, including: Obtaining a preset label embedding representation, wherein the label embedding representation has the same vector dimension as the high-order fusion label representation; Based on a preset tag scoring model, the tag score is calculated according to the tag embedding representation and the high-order fusion tag representation, and a user tag score is generated; Based on a preset tag filtering function, the tags are sorted and filtered according to the user tag scores, and a refined tag set is generated.

[0008] Specifically, user tags are obtained, and attention-weighted aggregation and graph convolution aggregation are performed on the user tags to generate a high-order fused tag representation, including: Obtain user tags, convert the user tag vector representation, and generate a tag vector set; Performing attention-weighted aggregation based on the tag vector set based on a preset attention-weighted aggregation model, and generating a preliminary fused tag representation; Graph convolution aggregation is performed based on the preliminary fusion label representation to generate a high-order fusion label representation.

[0009] Specifically, user tags are obtained, the user tag vector representation is converted, and a tag vector set is generated, including: Get user tag data; Extract the tag data source of user tag data; Setting a vector representation model according to the label data; Converting the user tag data into a tag vector representation according to the vector representation model; A label vector set is generated according to each of the label vector representations.

[0010] Specifically, based on a preset attention weighted aggregation model, attention weighted aggregation is performed according to the tag vector set, and a preliminary fusion tag representation is generated, including: Generate current feature vector based on user’s historical behavior data; Generate vector similarity and dynamic attention weight according to the label vector set and the current feature vector; Attention-weighted aggregation is performed according to the dynamic attention weights, and a preliminary fused label representation is generated.

[0011] Specifically, graph convolution aggregation is performed based on the preliminary fusion label representation to generate a high-order fusion label representation, including: Use the DeepWalk method to construct a graph structure between users or user tag representations; Each user is regarded as a node in the graph, and the cosine similarity of each user's preliminary fusion label representation is used as the edge between user nodes to determine the adjacency relationship and construct the adjacency matrix; The graph convolution propagation mechanism is used to further transform the preliminary fused label representation into a high-order fused representation.

[0012] Specifically, the self-supervised contrastive learning model includes: , in, is the contrast loss, which indicates the damage of the user's label vector in contrastive learning. The contrast loss is the optimization target. The smaller the value, the better the semantic aggregation effect. is a positive sample label set, which represents a user label set with high semantic similarity to the user. is the positive sample weight factor, is the user tag representation vector, which is the tag representation vector obtained based on the context-consistent tag set in step S300. is the label representation vector of another user with high semantic similarity to the user, which serves as a positive sample for contrastive learning. is a label representation in the comparison sample label set, that is, any label representation vector including positive samples and negative samples, is the similarity between the tags between users, which is calculated by cosine similarity. is the positive and negative difference coefficient, which is used to control the distribution difference between positive and negative samples in the contrast loss function and adjust the contrast intensity. is the label set of all comparison samples, including positive sample and negative sample label sets.

[0013] Specifically, the label scoring model includes: , in, Rating user tags, is the user's high-order fusion label representation vector, is the embedding representation vector of the jth label, that is, the semantic vector of the label, It is the similarity between the user's high-order fusion label representation vector and the label's semantic vector, which is used to measure the consistency between the user's interest and the label's semantics. It is obtained by calculating the cosine similarity. represents a hyperparameter that controls the contribution strength of the coupling enhancement term to the overall score. It is a coupling enhancement item used to measure the semantic synergy between a user's tag and all the user's historical tags, reflecting whether the tag is coordinated with the user's historical tag habits and whether it can form a semantically consistent tag combination with other tags. For the The embedding representation of the label, that is, the semantic vector of the label, For the jth label and the The relevance score of the tags is used to measure the semantic relevance between tags and is obtained by calculating the cosine similarity. is the coupling strength between labels, obtained through statistics or model training.

[0014] Specifically, generating vector similarity and dynamic attention weight according to the label vector set and the current feature vector includes: Based on the following formula, a vector similarity is generated according to the label vector set and the current feature vector: , in, is the label vector With the user's current feature vector The vector similarity between For the The label vector representation of the class label source, The user's current feature vector is generated based on user behavior data, historical records, or other methods, and contains the user's long-term interests and personalized preferences; Generate dynamic attention weights based on the following formula: , in, For the The dynamic attention weights related to the class label source represent the The importance of class labels to the final label representation of users is adjusted according to different users and scenarios. is the label vector With the user's current feature vector The vector similarity between is an exponential function.

[0015] Specifically, a user tag fusion system based on artificial intelligence is also provided, the system comprising: A high-order label representation generation module is used to obtain user labels, perform attention-weighted aggregation and graph convolution aggregation on the user labels, and generate a high-order fused label representation; A refined label set generation module, configured to perform multi-label sorting and classification on the high-order fusion label representation and generate a refined label set; A consistent tag set generation module, configured to perform user group co-occurrence clustering based on the generated refined tag set to generate a context consistent tag set; The optimized label set generation module is used to perform self-supervised contrastive learning based on the context-consistent label set based on a preset self-supervised contrastive learning model to obtain an optimized label set.

[0016] Specifically, the refined label set generation module is also used to: obtain a preset label embedding representation, wherein the label embedding representation has the same vector dimension as the high-order fusion label representation; based on a preset label scoring model, calculate the label score according to the label embedding representation and the high-order fusion label representation, and generate a user label score; based on a preset label filtering function, sort and filter the labels according to the user label score, and generate a refined label set.

[0017] Specifically, the high-order label representation generation module is also used to: obtain user labels, transform the user label vector representation, and generate a label vector set; perform attention-weighted aggregation based on the label vector set based on a preset attention-weighted aggregation model, and generate a preliminary fused label representation; perform graph convolution aggregation based on the preliminary fused label representation to generate a high-order fused label representation.

[0018] Specifically, the high-order label representation generation module is also used to: obtain user label data; extract the label data source of the user label data; set a vector representation model based on the label data; convert the user label data into a label vector representation according to the vector representation model; and generate a label vector set based on each of the label vector representations.

[0019] Specifically, the high-order label representation generation module is also used to: generate a current feature vector based on the user's historical behavior data; generate vector similarity and dynamic attention weight according to the label vector set and the current feature vector; perform attention weighted aggregation according to the dynamic attention weight, and generate a preliminary fused label representation.

[0020] Specifically, the high-order label representation generation module is also used to: use the DeepWalk method to construct a graph structure between users or user label representations, with each user as a node in the graph, and the cosine similarity of each user's preliminary fused label representation as the edge between user nodes to determine the adjacency relationship and construct an adjacency matrix; use the graph convolution propagation mechanism to further convert the preliminary fused label representation into a high-order fused representation.

[0021] Optionally, a computer device is also provided, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps described in the above-mentioned artificial intelligence-based user tag fusion method when executing the computer program.

[0022] Optionally, a computer-readable storage medium is also provided, on which a computer program is stored. When the computer program is executed by a processor, the steps described in the above-mentioned artificial intelligence-based user tag fusion method are implemented.

[0023] The present invention achieves the following technical effects: The above-mentioned artificial intelligence-based user tag fusion method and system utilizes machine learning and deep learning technologies to obtain user tags in turn, perform attention-weighted aggregation and graph convolution aggregation on the user tags, and generate a high-order fusion tag representation; perform multi-tag sorting and classification on the high-order fusion tag representation, and generate a refined tag set; perform user group co-occurrence clustering based on the generated refined tag set to generate a context-consistent tag set; based on a preset self-supervised comparative learning model, perform self-supervised comparative learning based on the context-consistent tag set to obtain an optimized tag set. Therefore, the present application enhances coupling between tags through a tag scoring model, takes into account the synergy and conflict relationships, improves the matching and correlation between users and tags, and introduces a tag sorting mechanism that can output tags in order of importance or relevance, thereby improving the quality of tags while ensuring the diversity of tags, thereby reducing the semantic redundancy of the tag set and reducing redundancy. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 1 is a flow chart of a method for user tag fusion based on artificial intelligence in one embodiment; Figure 2 This is a structural block diagram of an artificial intelligence-based user tag fusion system in one embodiment. DETAILED DESCRIPTION

[0025] In the following description, specific details such as specific system structures and techniques are provided for purposes of illustration rather than limitation to facilitate a thorough understanding of the embodiments of the present application. However, it will be apparent to those skilled in the art that the present application may be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to avoid obscuring the description of the present application with unnecessary detail.

[0026] It should be understood that when used in the present specification and the appended claims, the term "comprising" indicates the presence of described features, integers, steps, operations, elements and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or collections thereof.

[0027] It will also be understood that the term "and / or" used in this specification and the appended claims refers to and includes any and all possible combinations of one or more of the associated listed items.

[0028] As used in this specification and the appended claims, the term "if" can be interpreted as "when" or "upon" or "in response to determining" or "in response to detecting," depending on the context. Similarly, the phrase "if it is determined" or "if [described condition or event] is detected" can be interpreted as meaning "upon determination" or "in response to determining" or "upon detection of [described condition or event]" or "in response to detecting [described condition or event]," depending on the context.

[0029] In addition, in the description of the present application specification and the appended claims, the terms "first", "second", "third", etc. are only used to distinguish the descriptions and cannot be understood as indicating or implying relative importance.

[0030] References to "one embodiment" or "some embodiments" in this specification mean that a particular feature, structure, or characteristic described in conjunction with that embodiment is included in one or more embodiments of the present application. Thus, phrases such as "in one embodiment," "in some embodiments," "in other embodiments," and "in other embodiments" appearing in various places in this specification do not necessarily refer to the same embodiment, but rather mean "one or more but not all embodiments," unless otherwise specifically emphasized. The terms "including," "comprising," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0031] In one embodiment, a terminal is provided, which is used to: obtain user tags, perform attention-weighted aggregation and graph convolution aggregation on the user tags, and generate a high-order fused tag representation; perform multi-tag sorting and classification on the high-order fused tag representation, and generate a refined tag set; perform user group co-occurrence clustering based on the generated refined tag set to generate a context-consistent tag set; based on a preset self-supervised comparative learning model, perform self-supervised comparative learning on the context-consistent tag set to obtain an optimized tag set.

[0032] The terminal may be, but is not limited to, various personal computers, laptops, smart phones, tablet computers, and portable wearable devices.

[0033] In one embodiment, Figure 1 As shown, a user tag fusion method based on artificial intelligence is provided, and the method includes: Step S100: Obtain user tags, perform attention weighted aggregation and graph convolution aggregation on the user tags, and generate a high-order fused tag representation; Step S200: performing multi-label sorting and classification on the high-order fusion label representation, and generating a refined label set; Step S300: performing co-occurrence clustering of user groups based on the generated refined tag set to generate a context-consistent tag set; Step S400: Based on a preset self-supervised contrastive learning model, self-supervised contrastive learning is performed according to the context-consistent label set to obtain an optimized label set.

[0034] In this embodiment, by obtaining user tags, performing attention-weighted aggregation and graph convolution aggregation on the user tags, a high-order fused tag representation is generated; multi-tag sorting and classification is performed on the high-order fused tag representation, and a refined tag set is generated; user group co-occurrence clustering is performed based on the generated refined tag set to generate a context-consistent tag set; based on a preset self-supervised comparative learning model, self-supervised comparative learning is performed based on the context-consistent tag set to obtain an optimized tag set. Therefore, the present application enhances coupling between tags through a tag scoring model, takes into account collaborative and conflicting relationships, improves the matching and correlation between users and tags, and introduces a tag sorting mechanism that can output tags in order of importance or relevance, thereby improving the quality of tags while ensuring the diversity of tags, thereby reducing the semantic redundancy of the tag set and reducing redundancy.

[0035] In another embodiment, step S100: obtaining user tags, performing attention-weighted aggregation and graph convolution aggregation on the user tags, and generating a high-order fused tag representation includes: Step S110: obtaining user tags, converting the user tag vector representation, and generating a tag vector set; Step S120: performing attention weighted aggregation based on the tag vector set based on a preset attention weighted aggregation model, and generating a preliminary fused tag representation; Step S130: Perform graph convolution aggregation based on the preliminary fusion label representation to generate a high-order fusion label representation.

[0036] In this embodiment, in the AI-based user tag fusion method, user tag information, with user authorization, comes from a variety of sources, such as user behavioral data, content preferences, social relationships, and third-party data. Each tag source has its own unique characteristics and information dimensions, and these tags describe the user's interests, preferences, and behavior patterns in different scenarios. Each tag source has different formats, semantics, and expressions, and each type of tag describes different aspects of the user, including behavior, interests, social interactions, etc.

[0037] Therefore, it must be standardized into a consistent vector representation in order to perform unified processing, avoid information loss or misunderstanding, ensure that the semantic characteristics of each label are effectively retained, and convert different label sources into vector representations, which can effectively reduce the dimension of the data so that the data can be processed efficiently, thereby providing a unified basis for subsequent label fusion, greatly reducing the amount of calculation, improving the computational efficiency of the model, and improving the accuracy and stability in subsequent steps.

[0038] This step aims to convert user tag information from various sources, including behavior, content, social networks, and third parties, into a low-dimensional, structured vector representation, providing a unified and efficient input for subsequent tag fusion, modeling, and computation. Specifically, this involves mapping tags from various sources, such as behavior tags and content tags, into a common space. Using vector representation, the tag's semantic information, such as content and preferences, is embedded into a low-dimensional vector, preserving its inherent semantic structure. This vectorization reduces the dimensionality of the diverse data, enabling the model to perform computations more efficiently.

[0039] Then, based on the preset attention weighted aggregation model, attention weighted aggregation is performed according to the label vector set, and a preliminary fusion label representation is generated; graph convolution aggregation is performed based on the preliminary fusion label representation to generate a high-order fusion label representation.

[0040] In another embodiment, step S110: obtaining user tags, converting the user tag vector representation, and generating a tag vector set includes: Step S111: Obtain user tag data; Step S112: extracting the tag data source of the user tag data; Step S113: setting a vector representation model according to the label data; Step S114: converting the user tag data into a tag vector representation according to the vector representation model; Step S115: Generate a label vector set according to each of the label vector representations.

[0041] In this embodiment, the user tag sources in step S111 include behavior tags, content tags, social tags, and third-party tags.

[0042] The behavioral tags specifically include behavioral data such as searches, clicks, browsing, and purchases from users, which can reflect users' immediate interests and preferences. The content tags specifically include tags extracted based on users' interactions with content, such as comments and reading history, which can reflect users' long-term interests and values. The social tags specifically include tags generated from users' activities in social networks, such as following, liking, commenting, and other social behaviors, which can reveal users' social circles and social influence. The third-party tags specifically include tags provided by advertisers or other external data sources, such as demographic information, market research data, etc.

[0043] Selecting a label representation model involves choosing an appropriate label representation model for conversion, so that labels from different sources can be mapped into a unified space. For example, for text data, BERT is first used for pre-training, followed by fine-tuning based on the label content, so that the label semantics can be embedded in an efficient and rich vector space. The BERT model can handle the contextual relationships between labels, encoding each label as a low-dimensional vector while preserving its rich semantic information. For some behavioral and social labels, the Word2Vec model can be used to learn their semantic associations in user behavior. Word2Vec captures the semantic similarity between labels and generates a low-dimensional representation of each label. For some categorical labels (such as a user's gender or age), one-hot encoding can be used to convert the categorical information into a vector representation. Converting each label into a vector representation using this model preserves the label's core semantic information. The vector dimension of each label is typically hundreds of dimensions or higher to ensure sufficient representation power. Ultimately, each user's label is represented as a set of vectors, resulting in a label vector set.

[0044] In another embodiment, step S120: performing attention weighted aggregation based on the tag vector set based on a preset attention weighted aggregation model and generating a preliminary fused tag representation includes: Step S121: generating a current feature vector based on the user's historical behavior data; In this step, for each user, a personalized semantic feature vector, namely the current feature vector, is constructed based on their historical behavior logs, browsing content, interest evolution, etc., to reflect the user's overall preference trend. Combined with the label vector, the vector similarity and attention weight are calculated.

[0045] Step S122: generating vector similarity and dynamic attention weight according to the label vector set and the current feature vector; Step S123: Perform attention weighted aggregation according to the dynamic attention weights and generate a preliminary fusion label representation.

[0046] In this embodiment, after obtaining the label vector set, each label source in the label vector set represents a different dimension of user behavior, and the relative importance of these label sources will vary according to the actual application scenario. Therefore, it is necessary to reasonably weight these labels from different sources to improve the accuracy and relevance of label fusion, and enable the aggregated label representation to more accurately reflect the user's interests and preferences. This step uses an attention-weighted aggregation model based on the obtained label vector set, dynamically adjusts the weights of different label sources through the semantic content of the label and the current context, performs attention-weighted aggregation on the label vector, and obtains a preliminary fused label representation. Specifically, vector similarity and dynamic attention weights are generated based on the label vector set and the current feature vector; attention-weighted aggregation is then performed based on the dynamic attention weight, and a preliminary fused label representation is generated.

[0047] In another embodiment, step S122: generating vector similarity and dynamic attention weight according to the tag vector set and the current feature vector, includes: Step S1221: Generate vector similarity based on the label vector set and the current feature vector based on the following formula: , in, is the label vector With the user's current feature vector The vector similarity between For the The label vector representation of the class label source, The user's current feature vector is generated based on user behavior data, historical records, or other methods, and contains the user's long-term interests and personalized preferences; Step S1222: Generate dynamic attention weight based on the following formula: , in, For the The dynamic attention weights related to the class label source represent the The importance of class labels to the final label representation of users is adjusted according to different users and scenarios. is the label vector With the user's current feature vector The vector similarity between is an exponential function.

[0048] Furthermore, in step S123, an attention-weighted aggregation model is used to dynamically adjust the weights of different tag sources based on the semantic content of the tag and the current context, and perform attention-weighted aggregation on the tag vector to obtain a preliminary fused tag representation. The attention-weighted aggregation model includes: ,in, is the initial fusion label representation, that is, the label vector representation obtained after the user's labels are initially fused. is the total number of tag sources, For the Dynamic attention weights related to class label sources, For the Label vector representation of class label sources, including behavioral labels, content labels, social labels, and third-party labels.

[0049] Therefore, considering that traditional label weighting methods usually use fixed weights to weight different label sources, there is a problem that they cannot accurately reflect the user's true preferences at a specific time or in a specific situation, and the user's preferences and behaviors are constantly changing, the traditional fixed weight method cannot adapt to such changes. However, this model can adaptively adjust the importance of different label sources by dynamically calculating the attention weights, and weight different labels according to the user's personalized characteristics (such as historical behavior, changes in interests, etc.), thereby improving the adaptability of the model, improving the accuracy of label fusion, and enhancing the flexibility of the model.

[0050] In another embodiment, step S130: performing graph convolution aggregation based on the preliminary fused label representation to generate a high-order fused label representation includes: Step S131: using the DeepWalk method to construct a graph structure between users or user tag representations; Step S132: Taking each user as a node in the graph, and using the cosine similarity of each user's preliminary fusion label representation as the edge between user nodes to determine the adjacency relationship and construct an adjacency matrix; Step S133: Utilize the graph convolution propagation mechanism to further convert the preliminary fused label representation into a high-order fused representation.

[0051] In this step, after obtaining the preliminary fusion label representation vector, the vector has integrated information from multiple label sources and has a certain expressive ability. However, user interests often show group trends, such as being affected by hot topics and social interactions, and there are interest diffusion and label influence paths. However, this representation is still a fusion in the local dimension, that is, the fusion process only fuses individual label sources within the single user dimension, and does not consider the potential correlation between users or the structural relationship of labels at the group level. Therefore, in order to further extract the high-order relationship between user groups in the label semantic structure, it is necessary to introduce graph structure modeling and graph convolution aggregation mechanism to perform high-order semantic modeling on the label representation between users, thereby strengthening the semantic association and propagation path modeling in the preliminary fusion label, generating a high-order label fusion representation with global semantic context, enhancing the model's adaptability to cold-start users or label-sparse scenarios, and improving the model's stability and robustness.

[0052] Furthermore, the DeepWalk method is first used to construct a graph structure between user or user label representations. Each user is treated as a node in the graph, and the cosine similarity of each user's preliminary fused label is used as the edge between user nodes to determine the adjacency relationship. An adjacency matrix is constructed, where the value is 1 if two users are connected and 0 otherwise. Then, using the graph convolutional propagation mechanism and GCN (Graph Convolutional Network) technology, a standard two-layer graph convolution operation is used to further transform the preliminary fused label representation into a high-order fused representation.

[0053] The final output is a high-order fusion label representation, which integrates the user's preliminary fusion label and the semantic information of neighboring users. It has the ability to propagate graph structures and can be used to build cross-user label relationships, significantly improving the semantic integrity of label expression.

[0054] In another embodiment, step S200: performing multi-label sorting and classification on the high-order fusion label representation and generating a refined label set includes: Step S210: obtaining a preset label embedding representation, wherein the label embedding representation has the same vector dimension as the high-order fusion label representation; In this step, a label embedding representation is initialized for each label , this embedding representation converts the semantic space of the label into a vector form, and is combined with the high-order fusion label representation The vectors have the same dimensions.

[0055] Step S220: Based on a preset tag scoring model, the tag score is calculated according to the tag embedding representation and the high-order fusion tag representation, and a user tag score is generated; In this step, the tag score is calculated based on the initialized tag embedding representation and the high-order fusion tag representation to obtain the user tag score ,The score reflects the similarity between the user and the tag in the ,semantic space, which helps to judge the user’s potential interest in the tag and ,provide more accurate tag recommendations.

[0056] Step S230: sorting and filtering the tags according to the user tag scores based on a preset tag filtering function, and generating a refined tag set.

[0057] In this step, the user tag score is calculated , use the label filter function , sort and filter the tags, and output the top tags with the highest scores and the most relevant to the user. tags and combine them to obtain a refined tag set, thereby ensuring the relevance and diversity of tags and avoiding the interference of tag redundancy and low-quality tags.

[0058] In this embodiment, the vector represented by the obtained high-order fusion label represents the fusion of the semantic relationship between the label sources and the group structure relationship between users. However, the vector itself is still a continuous vector representation. The label sources are diverse, and there are problems such as redundant labels and semantic repetition. It does not have clear label semantics and cannot be directly used. In addition, user preferences are often not single, but a combination of multiple labels. The correlation and sorting priority between labels must be considered.

[0059] Therefore, it is necessary to map the high-order fusion label representation vector into a discrete label set with actual label semantics, and screen out a representative and refined label set that best expresses user interests through a sorting and classification mechanism, so as to reflect user interests and control label redundancy.

[0060] This step uses a tag scoring model based on the obtained high-order fusion tag representation to score the tags based on the degree of consistency between user interests and tag semantics, combined with the synergy and conflict relationships between tags. The tags are then sorted and output based on the scoring results, thereby obtaining the most representative tag set for the user, namely the refined tag set, which improves the refinement, accuracy and semantic consistency of the tag set and reduces the number of irrelevant tags.

[0061] In another embodiment, the tag scoring model includes: , in, Rating user tags, is the user's high-order fusion label representation vector, is the embedding representation vector of the jth label, that is, the semantic vector of the label, It is the similarity between the user's high-order fusion label representation vector and the label's semantic vector, which is used to measure the consistency between the user's interest and the label's semantics. It is obtained by calculating the cosine similarity. represents a hyperparameter that controls the contribution strength of the coupling enhancement term to the overall score. It is a coupling enhancement item used to measure the semantic synergy between a user's tag and all the user's historical tags, reflecting whether the tag is coordinated with the user's historical tag habits and whether it can form a semantically consistent tag combination with other tags. For the The embedding representation of the label, that is, the semantic vector of the label, For the jth label and the The relevance score of the tags is used to measure the semantic relevance between tags and is obtained by calculating the cosine similarity. is the coupling strength between labels, obtained through statistics or model training.

[0062] Specifically, ,in, represents the product of the Euclidean lengths of two vectors, represents the inner product of two vectors. Similarly, The calculation method is the same as that of , and those skilled in the art should know how to calculate, so this application will not go into details.

[0063] For the inter-tag coupling strength , ranging from 0 to 1. For example, according to statistical results, users with the "basketball" tag are very likely to also have the "sports" tag, that is, the "sports" tag is highly correlated with the "basketball" tag, and the coupling strength between the two tags is high, close to 1.

[0064] The label sorting output includes: , in, To refine the tag set, filter the output of the most relevant tags to the user. A collection of tags, It is a tag filtering function that sorts tags according to user tag scores and selects the top tags with the highest scores. labels, and C represents the total number of candidate labels.

[0065] The above model is used to perform multi-label sorting and classification on the high-order fusion label representation to obtain refined labels.

[0066] Traditional multi-label ranking and classification methods usually use threshold screening, linear classifiers or label-by-label probability regression for multi-label recognition. These methods ignore the ranking dependency between labels, ignore the mutual exclusivity or joint probability between labels, reduce the correlation between users and labels, and lack a label granularity control mechanism, resulting in high redundancy and low quality of the obtained labels. This model enhances coupling between labels, considers synergy and conflict relationships, improves the matching and correlation between users and labels, and introduces a label sorting mechanism that can output labels in order of importance or relevance, improving label quality while ensuring label diversity.

[0067] In another embodiment, in step S300: performing user group co-occurrence clustering based on the generated refined tag set to generate a context-consistent tag set, the user group co-occurrence clustering is specifically as follows: The refined tag set obtained, the tags in this set reflect the user's interests and behavioral characteristics. However, the tag set of a single user may be affected by noise and cannot fully represent the user's behavior pattern in a certain group. In order to better capture the contextual consistency between user tags, that is, the co-occurrence of user tags in a group, this step further optimizes the tags through the group co-occurrence clustering method, aggregates user group tags, and identifies the co-occurrence patterns between user tags based on user behavior data at the group level, further improves the contextual consistency of the tag set, optimizes tag relevance, eliminates possible noise and deviations in single user tags, enhances the representativeness of the tag set in similar user groups, improves the group recommendation effect, and constructs a tag set that can better reflect the group's interests through group behavior analysis, provides a more accurate tag set for the group recommendation system, and finally obtains a contextually consistent tag set.

[0068] Based on the obtained refined tag set, the steps of group co-occurrence clustering include the construction of a tag co-occurrence matrix, the calculation of a weighted co-occurrence matrix, and graph clustering. The construction of a tag co-occurrence matrix refers to the use of the number of times tag one and tag two co-occur among all users to form a tag co-occurrence matrix in a matrix with a dimension of the number of tags × the number of tags. This step reveals the relative correlation between tags. For example, the tags "sports" and "basketball" may co-appear in many users. In order to avoid the direct influence of the frequency of tag co-occurrence, the tags are weighted according to their frequency of occurrence among all users, and a weighted co-occurrence matrix is calculated to reflect the prevalence of the tags. This step adjusts the co-occurrence frequency through the tag weights to ensure that tags that appear frequently and have a high prevalence have more influence in the cluster. Finally, graph clustering is performed. The weighted co-occurrence matrix is used to cluster the labels using the spectral clustering algorithm. The clustering results of the labels are obtained by calculating the eigenvalues and eigenvectors of the Laplacian matrix of the graph. This clustering process divides the labels into multiple clusters. Each cluster represents a set of semantically related labels that frequently co-occur in the group, that is, a context-consistent label set. This step can capture the group co-occurrence pattern between labels and identify the contextual correlation between labels.

[0069] In another embodiment, in step S400, based on a preset self-supervised contrastive learning model, self-supervised contrastive learning is performed according to the context-consistent label set to obtain an optimized label set.

[0070] The context-consistent tag set is pre-acquired. These tags have been semantically clustered based on co-occurrence relationships across user groups. However, there may still be semantic overlap between different users, where similar tags are distributed but express different intentions. Therefore, it is necessary to introduce a self-supervised contrastive learning mechanism to further structurally align and differentiate user tag representations, improve the generalization of tag representations and the discriminant boundaries between different semantic tags, and enable the model to more accurately distinguish similar tag sets, thereby improving the discriminability of tag vectors, enhancing the aggregation and dispersion of tag representations in the embedding space, and strengthening the discriminability and semantic consistency of tag representations.

[0071] Based on the obtained context-consistent label set, a self-supervised contrastive learning model is used to calculate the contrast loss with the positive and negative sample sets. By minimizing the contrast damage, the discriminant enhancement of the user label representation space is achieved, and the aggregation and discriminability of the label representation in the multi-user semantic space are improved. The optimized label embedding representation vector is obtained, and the context-consistent label set is optimized to obtain the optimized label set, thereby improving the accuracy and robustness of label fusion and the recognition of labels.

[0072] In another embodiment, the self-supervised contrastive learning model includes: , in, is the contrast loss, which means the label vector of user i represents the damage size in contrastive learning. Contrastive loss is the optimization target. The smaller the value, the better the semantic aggregation effect. i is the index of the user, indicating which user the label calculation is performed for. The range of i is all users used for model training. is a positive sample label set, which represents a user label set with high semantic similarity to the user. is the positive sample weight factor, is the user tag representation vector, which is the tag representation vector obtained based on the context-consistent tag set in step S300. is the label representation vector of another user with high semantic similarity to the user, which serves as a positive sample for contrastive learning. is a label representation in the comparison sample label set, that is, any label representation vector including positive samples and negative samples, is the similarity between the tags between users, which is calculated by cosine similarity. is the positive and negative difference coefficient, which is used to control the distribution difference between positive and negative samples in the contrast loss function and adjust the contrast intensity. is the label set of all comparison samples, including positive sample and negative sample label sets.

[0073] Traditional contrastive learning relies on a large number of negative samples, which may result in the semantic differences between labels not being particularly obvious, thus affecting the learning effect of the model. In addition, traditional contrastive learning does not fully consider the complex coupling relationship between labels and tends to ignore the contextual consistency of the label set.

[0074] This model uses semantic consistency between labels to weight the comparison between positive and negative samples, so that label learning no longer depends solely on the number of negative samples, thereby reducing the noise introduced by negative samples and improving the accuracy of label representation. By considering the contextual consistency of labels, our model can more reasonably aggregate labels in different contexts and avoid label confusion.

[0075] Furthermore, the steps are as follows: First, generate comparison samples: select users who have some of the same label clusters or similar clustering features as the user to form a positive sample set , and set the positive sample weight factor according to the label overlap , randomly sample from users with large differences in user labels to form a negative sample set, and combine the positive sample set with the negative sample set to obtain the set of all comparison sample labels ; Secondly, loss function construction: construct a multi-granularity weighted contrast loss function, perform contrast learning on positive and negative samples for each user label representation vector, and calculate contrast loss , and use the positive and negative coefficient of variation Control the sensitivity of sample distribution in the loss function and enhance the expressive power; Finally, model training: Train the model to minimize the contrast loss of all users, achieve discriminant enhancement in the user label representation space, improve the aggregation and structure of the label representation vector in the semantic space, optimize the label representation vector so that similar labels express the same semantics, improve the semantic consistency of the label representation vector, and solve the problem of label semantic aliasing in traditional representation learning.

[0076] The following is an example of an application scenario of the present application: for example, on an e-commerce platform, the platform obtains a set of labels, including "sports shoes", "running shoes" and "fitness equipment". These labels are obtained by co-occurrence clustering of user purchasing behaviors. On the surface, they are semantically similar, but different users have subtle differences in their understanding and interests of these labels. User A focuses on comfort and durability, user B focuses on appearance, and user C focuses on effects. These user labels may be confused after co-occurrence clustering, and the semantic representation may be vague, making it difficult to accurately reflect the user's true intentions. Therefore, through self-supervised comparative learning in step S6, the discriminability and semantic consistency of the label representation are enhanced to form three new label sets with obvious semantic distinctions, such as "comfortable wear, durable equipment" for user A, "trendy fashion, fashionable shoes" for user B, and "sports health, fitness equipment" for user C.

[0077] In one embodiment, Figure 2 As shown, a user tag fusion system based on artificial intelligence is also provided, and the system includes: A high-order label representation generation module is used to obtain user labels, perform attention-weighted aggregation and graph convolution aggregation on the user labels, and generate a high-order fused label representation; A refined label set generation module, configured to perform multi-label sorting and classification on the high-order fusion label representation and generate a refined label set; A consistent tag set generation module, configured to perform user group co-occurrence clustering based on the generated refined tag set to generate a context consistent tag set; The optimized label set generation module is used to perform self-supervised contrastive learning based on the context-consistent label set based on a preset self-supervised contrastive learning model to obtain an optimized label set.

[0078] In another embodiment, the refined label set generation module is further used to: obtain a preset label embedding representation, wherein the label embedding representation has the same vector dimension as the high-order fusion label representation; based on a preset label scoring model, calculate the label score according to the label embedding representation and the high-order fusion label representation, and generate a user label score; based on a preset label filtering function, sort and filter the labels according to the user label score, and generate a refined label set.

[0079] In another embodiment, the high-order label representation generation module is also used to: obtain user labels, transform the user label vector representation, and generate a label vector set; perform attention-weighted aggregation on the label vector set based on a preset attention-weighted aggregation model, and generate a preliminary fused label representation; perform graph convolution aggregation based on the preliminary fused label representation to generate a high-order fused label representation.

[0080] In another embodiment, the high-order label representation generation module is also used to: obtain user label data; extract the label data source of the user label data; set a vector representation model based on the label data; convert the user label data into a label vector representation based on the vector representation model; and generate a label vector set based on each of the label vector representations.

[0081] In another embodiment, the high-order label representation generation module is also used to: generate a current feature vector based on the user's historical behavior data; generate vector similarity and dynamic attention weights based on the label vector set and the current feature vector; perform attention weighted aggregation based on the dynamic attention weights, and generate a preliminary fused label representation.

[0082] In another embodiment, the high-order label representation generation module is also used to: use the DeepWalk method to construct a graph structure between users or user label representations, with each user as a node in the graph, and the cosine similarity of each user's preliminary fused label representation as the edge between user nodes to determine the adjacency relationship and construct an adjacency matrix; use the graph convolution propagation mechanism to further convert the preliminary fused label representation into a high-order fused representation.

[0083] In one embodiment, a computer device is provided, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps described in the above-mentioned artificial intelligence-based user tag fusion method when executing the computer program.

[0084] In one embodiment, a computer-readable storage medium is further provided, on which a computer program is stored. When the computer program is executed by a processor, the steps of the above-mentioned artificial intelligence-based user tag fusion method are implemented.

[0085] It should be noted that the information interaction, execution process and other contents between the above modules are based on the same concept as the method embodiment of this application. Their specific functions and technical effects can be found in the method embodiment part and will not be repeated here.

[0086] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated into one processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0087] It should be noted that the information interaction, execution process and other contents between the above modules are based on the same concept as the method embodiment of this application. Their specific functions and technical effects can be found in the method embodiment part and will not be repeated here.

[0088] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated into one processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0089] An embodiment of the present application also provides a network device, which includes: at least one processor, a memory, and a computer program stored in the memory and executable on the at least one processor, wherein the processor implements the steps of any of the above-mentioned method embodiments when executing the computer program.

[0090] An embodiment of the present application further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps in the above-mentioned various method embodiments can be implemented.

[0091] An embodiment of the present application provides a computer program product. When the computer program product is run on a mobile terminal, the mobile terminal can implement the steps in the above-mentioned various method embodiments when executing the computer program product.

[0092] If the integrated unit is implemented as a software functional unit and sold or used as a standalone product, it can be stored in a computer-readable storage medium. Based on this understanding, the present application implements all or part of the process steps in the above-mentioned method embodiments by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When executed by a processor, the computer program can implement the steps of each of the above-mentioned method embodiments. The computer program includes computer program code, which can be in source code form, object code form, executable file, or some intermediate form. The computer-readable medium can include at least: any entity or device capable of carrying computer program code to a camera / terminal device, recording medium, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signals, telecommunication signals, and software distribution media. Examples include USB flash drives, removable hard drives, magnetic disks, or optical disks. In some jurisdictions, based on legislation and patent practice, computer-readable media cannot be electric carrier signals or telecommunication signals.

[0093] In the above embodiments, the description of each embodiment has its own focus. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant description of other embodiments.

[0094] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0095] In the embodiments provided in this application, it should be understood that the disclosed devices / network equipment and methods can be implemented in other ways. For example, the device / network equipment embodiments described above are merely illustrative. For example, the division of the modules or units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0096] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0097] The above-described embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present application, and should all be included in the scope of protection of the present application.

[0098] An embodiment of the present application also provides a computer device, which includes: at least one processor, a memory, and a computer program stored in the memory and executable on the at least one processor, wherein the processor implements the steps in any embodiment of the above method when executing the computer program.

[0099] The computer device may include, but is not limited to, a processor and a memory. Those skilled in the art will appreciate that the above description is an example of a computer device and does not limit the computer device. The computer device may include more or fewer components than described above, or a combination of certain components, or different components. For example, the computer device may also include input / output devices, network access devices, etc.

[0100] The processor may be a central processing unit (CPU), and the processor 0 may also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0101] In some embodiments, the memory may be an internal storage unit of the computer device, such as a hard disk or memory of the computer device. In other embodiments, the memory may also be an external storage device of the computer device, such as a plug-in hard disk, a Smart Media Card (SMC), a Secure Digital (SD) card, a Flash Card, etc. equipped with the computer device. Furthermore, the memory may include both an internal storage unit of the computer device and an external storage device. The memory is used to store an operating system, application programs, a boot loader, data, and other programs, such as the program code of the computer program. The memory may also be used to temporarily store data that has been output or is about to be output.

[0102] 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.

[0103] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present application, and such modifications and improvements are intended to fall within the scope of protection of the present application. Therefore, the scope of protection of the present application shall be determined by the appended claims.

Claims

1. A user tag fusion method based on artificial intelligence, characterized in that: The method comprises: Obtain user tags, perform attention-weighted aggregation and graph convolution aggregation on the user tags, and generate a high-order fused tag representation; Performing multi-label sorting and classification on the high-order fusion label representation and generating a refined label set; Performing co-occurrence clustering of user groups based on the generated refined tag set to generate a context-consistent tag set; Based on a preset self-supervised contrastive learning model, self-supervised contrastive learning is performed according to the context-consistent label set to obtain an optimized label set.

2. The user tag fusion method based on artificial intelligence according to claim 1 is characterized in that: Perform multi-label sorting and classification on the high-order fusion label representation and generate a refined label set, including: Obtaining a preset label embedding representation, wherein the label embedding representation has the same vector dimension as the high-order fusion label representation; Based on a preset tag scoring model, the tag score is calculated according to the tag embedding representation and the high-order fusion tag representation, and a user tag score is generated; Based on a preset tag filtering function, the tags are sorted and filtered according to the user tag scores, and a refined tag set is generated.

3. The user tag fusion method based on artificial intelligence according to claim 1 is characterized in that: Obtain user tags, perform attention-weighted aggregation and graph convolution aggregation on the user tags, and generate a high-order fused tag representation, including: Obtain user tags, convert the user tag vector representation, and generate a tag vector set; Performing attention-weighted aggregation based on the tag vector set based on a preset attention-weighted aggregation model, and generating a preliminary fused tag representation; Graph convolution aggregation is performed based on the preliminary fusion label representation to generate a high-order fusion label representation.

4. The user tag fusion method based on artificial intelligence according to claim 3 is characterized in that: Obtain user tags, transform the user tag vector representation, and generate a tag vector set, including: Get user tag data; Extract the tag data source of user tag data; Setting a vector representation model according to the label data; Converting the user tag data into a tag vector representation according to the vector representation model; A label vector set is generated according to each of the label vector representations.

5. The user tag fusion method based on artificial intelligence according to claim 3 is characterized in that: Performing attention-weighted aggregation based on the tag vector set based on a preset attention-weighted aggregation model and generating a preliminary fused tag representation, including: Generate current feature vector based on user’s historical behavior data; Generate vector similarity and dynamic attention weight according to the label vector set and the current feature vector; Attention-weighted aggregation is performed according to the dynamic attention weights, and a preliminary fused label representation is generated.

6. The user tag fusion method based on artificial intelligence according to claim 3 is characterized in that: Perform graph convolution aggregation based on the preliminary fusion label representation to generate a high-order fusion label representation, including: Use the DeepWalk method to construct a graph structure between users or user tag representations; Each user is regarded as a node in the graph, and the cosine similarity of each user's preliminary fusion label representation is used as the edge between user nodes to determine the adjacency relationship and construct the adjacency matrix; The graph convolution propagation mechanism is used to further transform the preliminary fused label representation into a high-order fused representation.

7. The user tag fusion method based on artificial intelligence according to claim 1 is characterized in that: The self-supervised contrastive learning model includes: ， in, is the contrast loss, is the positive sample label set, is the positive sample weight factor, is the user label representation vector, i is the index of the user label, is the label representation vector of another user with high semantic similarity to this user, is a label representation vector in the comparison sample label set, Represents the user label vector With label representation vector The similarity between Represents the user label representation vector Vector with labels The similarity between labels, is the positive and negative coefficient of variation, is the label set of all comparison samples, including positive sample and negative sample label sets.

8. The user tag fusion method based on artificial intelligence according to claim 2 is characterized in that: The label scoring model includes: ， in, Rating user tags, is the user's high-order fusion label representation vector, is the embedding representation vector of the jth label, Represents the set of all historical tags that a user has. is the similarity between the user’s high-order fusion label representation vector and the label’s semantic vector, represents the hyperparameter, For the The embedding representation of the labels, is the jth label and the The relevance score of the tags, is a coupling enhancement term used to measure the semantic synergy between user tags and all user historical tags. is the coupling strength between tags.

9. The user tag fusion method based on artificial intelligence according to claim 5 is characterized in that: Generating a vector similarity and a dynamic attention weight according to the label vector set and the current feature vector, including: Based on the following formula, a vector similarity is generated according to the label vector set and the current feature vector: ， in, is the label vector With the user's current feature vector The vector similarity between For the The label vector representation of the class label source, is the user's current feature vector; Generate dynamic attention weights based on the following formula: ， in, For the The dynamic attention weights related to the class label source represent the The importance of class labels to the user's final label representation, is the label vector With the user's current feature vector The vector similarity between is an exponential function, The total number of tag sources.

10. A user tag fusion system based on artificial intelligence, characterized in that: The system comprises: A high-order label representation generation module is used to obtain user labels, perform attention-weighted aggregation and graph convolution aggregation on the user labels, and generate a high-order fused label representation; A refined label set generation module, configured to perform multi-label sorting and classification on the high-order fusion label representation and generate a refined label set; A consistent tag set generation module, configured to perform user group co-occurrence clustering based on the generated refined tag set to generate a context consistent tag set; The optimized label set generation module is used to perform self-supervised contrastive learning based on the context-consistent label set based on a preset self-supervised contrastive learning model to obtain an optimized label set.

Images