A blockchain-based micro-creative resource semantic understanding and intelligent recommendation method

Abstract

The application discloses a kind of micro-creative resource semantic understanding and intelligent recommendation method based on blockchain.Collect the multi-modal content of micro-creative resource and its blockchain history transmission path, generate initial embedding vector;Based on user continuous attention trajectory, construct transmission directed graph, by graph attention mechanism, the transmission behavior characteristics is fused into vertex representation, and graph enhanced micro-creative resource vector is obtained;Cross-modal positive sample pair is constructed using user continuous attention trajectory, and graph enhanced micro-creative resource vector is mapped to uniform semantic space;In uniform semantic space, introduce transmission influence constraint to carry out model training, and add transmission influence prediction branch;Receive query, after mapping to uniform semantic space, carry out similarity matching with candidate micro-creative resource, and obtain recommendation result in combination with predicted influence.The application realizes the intelligent recommendation of integrated semantic matching and influence evaluation.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more particularly to a blockchain-based method for semantic understanding and intelligent recommendation of micro-creative resources. This invention also relates to systems, electronic devices, and storage media for implementing the above method. Background Technology

[0002] With the rapid development of blockchain technology, micro-creative resources, as an emerging form of multimodal creative content, are gradually attracting attention for their creation and dissemination. Micro-creative resources typically encompass creative content in multiple modalities, including text, images, and audio. Their dissemination process is recorded on the blockchain, leaving an immutable record containing rich information about dissemination behavior, such as user account identifiers, attention timestamps, and dissemination popularity weights. However, how to effectively perform semantic understanding and intelligent recommendation of these micro-creative resources, enabling precise matching at the content level while also reflecting the dissemination influence and path characteristics reflected in the blockchain records, has become a core problem that urgently needs to be solved in the current technological system.

[0003] The existing technology has the following shortcomings:

[0004] 1. Existing micro-creative resource modeling ignores the propagation structure information in blockchain records: Conventional methods only process content data such as text, images, and audio, without utilizing the continuous attention trajectory of user accounts, the diffusion effect of recently popular content, and the group behavior signals reflected by repeated attention. As a result, micro-creative resource representation lacks a propagation ecosystem dimension, and the recommendation results can only reflect content similarity, making it difficult to reflect the actual attention and potential propagation influence of micro-creative resources.

[0005] 2. Existing cross-modal alignment training relies on manual annotation, which cannot adapt to the dynamic changes in multimodal associations: Associations between different modal micro-creative resources (such as text and images, text and audio) usually require manual annotation one by one, which is costly and difficult to cover the constantly emerging new micro-creative resources and new modal combinations. Existing methods lack the ability to discover cross-modal associations from real user attention behavior, making it difficult to effectively compare and recommend multimodal micro-creative resources in the same semantic space.

[0006] 3. Existing semantic spaces lack the ability to represent the gradient of dissemination influence: Traditional semantic models are trained only to bring similar content closer together and push away irrelevant content, without distinguishing between micro-creative resources with different dissemination influences under the same topic. This results in high-influence and low-influence micro-creative resources being mixed together in the semantic space, failing to provide effective reference for users' content discovery needs (such as prioritizing micro-creative resources with higher creative reference value), and also reducing the reasonableness of the content quality of the recommendation results.

[0007] 4. Existing recommendation methods separate semantic matching from influence ranking: A common approach is to first recall candidate micro-creative resources based on content similarity, and then rank them separately according to historical popularity or other influence indicators. This two-stage approach easily leads to recommendation results with high semantic matching but low dissemination influence. Furthermore, it lacks effective influence assessment and ranking criteria for cold-start micro-creative resources with limited attention records, making it difficult to meet the intelligent recommendation needs of micro-creative resource recommendation platforms that require both relevance and dissemination influence.

[0008] Therefore, there is an urgent need for an intelligent recommendation method for micro-creative resources that can deeply integrate user attention behavior recorded in the blockchain into multimodal micro-creative resource representation learning, mine cross-modal associations, construct a unified semantic space for perceiving the gradient of dissemination influence, and achieve integrated and coordinated semantic matching and influence ranking, in order to solve the above problems. Summary of the Invention

[0009] This invention addresses the technical problems in existing technologies, such as micro-creative resource modeling ignoring propagation structure information in blockchain records, cross-modal alignment relying on manual annotation leading to high costs and difficulty in adaptation, semantic space lacking the ability to express propagation influence gradients, and recommendation methods separating semantic matching and influence ranking resulting in low propagation influence of recommendation results and lack of evaluation basis for cold-start micro-creative resources. It provides a blockchain-based method for semantic understanding and intelligent recommendation of micro-creative resources.

[0010] Therefore, the first objective of this invention is to provide a blockchain-based method for semantic understanding and intelligent recommendation of micro-creative resources, which includes the following steps: Acquire multimodal content data from micro-creative resources to generate initial embedding vectors; Based on the historical propagation path of the micro-creative resources on the blockchain, a propagation directed graph is constructed with the micro-creative resources as vertices, and the initial embedding vector is used as the initial feature. The propagation behavior features are integrated into the vertex representation through a graph neural network to obtain the graph-enhanced micro-creative resource vector. By contrastive learning, the graph-enhanced micro-creative resource vectors are mapped to a unified semantic space, and a propagation influence constraint is introduced into the unified semantic space for model training. At the same time, a propagation influence prediction branch is added to obtain the predicted influence of the micro-creative resources. The system receives queries, encodes them, maps them to the unified semantic space to obtain a query representation, selects candidate resources from the micro-creative resources, performs similarity matching between the query representation and the representation of the candidate resources, and combines this with the predicted influence to obtain a recommendation result.

[0011] Preferably, generating the initial embedding vector includes: Each modality in the multimodal content data is encoded separately to obtain the initial features of each modality; Construct a modal effective vector, wherein each component of the modal effective vector corresponds one-to-one with each mode; The initial features of each modality are input into the modality projection layer corresponding to that modality to obtain the modality projection features; Based on the modality effective vector, the gating score corresponding to the missing modality is assigned a minimum value, and the gating score of the existing modality is normalized to obtain the gating weight. The initial embedding vector is obtained by weighting and summing the projection features of each modality according to the gating weights.

[0012] Preferably, the construction of the propagation directed graph and the integration of propagation behavior features into the vertex representation through a graph neural network to obtain a graph-enhanced micro-creative resource vector include: Using the micro-creative resources as vertices, the historical propagation paths are sorted chronologically by user accounts. Directed edges are established between adjacent vertices and the number of repeated attentions is accumulated. Edge weights are calculated based on the number of repeated attentions, the propagation popularity weights in the historical propagation paths, and the attention time intervals in the historical propagation paths. Self-loop edges are added and adjacent vertices whose weights meet preset conditions are retained. Construct a feature vector containing the edge weights and the directions of the directed edges, map it to a behavior bias and superimpose it on the content relevance score, normalize it to obtain an attention coefficient, and aggregate the neighbor vertex features according to the weight of the coefficient to update the center vertex features, thus obtaining a graph-enhanced micro-creative resource vector.

[0013] Preferably, the step of mapping graph-enhanced micro-creative resource vectors to a unified semantic space through contrastive learning includes: The graph-enhanced micro-creative resource vectors are input into the shared projection layer to obtain the shared basic vectors; Select the corresponding modality-specific projection module based on the main modality of the micro-creative resource, and output the modality compensation vector; The shared fundamental vector and the modal compensation vector are added together, and the result is layer-normalized to obtain the representation in the unified semantic space. Based on the modality effective vector, the main modality of each micro-creative resource is determined, and pairs of micro-creative resources that are followed sequentially within a preset time window and have different main modalities are identified as cross-modality positive sample pairs. Using the cross-modal positive sample pairs as positive samples for contrastive learning, the two representations corresponding to the cross-modal positive sample pairs are brought closer together in the unified semantic space.

[0014] Preferably, the step of introducing a propagation influence constraint into the unified semantic space for model training, while simultaneously adding a propagation influence prediction branch to obtain the predicted influence of the micro-creative resource, includes: Micro-creative resources are categorized into high, medium, and low-impact sets based on their dissemination popularity weight; Anchor points are selected from high or medium influence sets, positive samples are selected from resources with the same or higher influence level as the anchor points, and negative samples are selected from resources with lower influence levels than the anchor points to form triples. The contrast loss margin is adjusted based on the influence difference of the triples, and a compact regularization constraint is applied to the representations in the high-influence set, wherein the margin comprises the product of the base margin and the dynamic adjustment term. By propagating the influence prediction branch, the influence is predicted based on the representation in the unified semantic space.

[0015] Preferably, the method of obtaining the recommendation result by combining the predicted influence includes: Calculate the cosine similarity between the query representation and the candidate resource representation; A recommendation score is calculated based on the cosine similarity and the predicted influence of the candidate resource. The recommendation score is the sum of the cosine similarity and a first product, where the first product is the product of the cosine similarity, a preset coefficient, and the predicted influence of the candidate resource. Candidate resources are sorted from highest to lowest according to their recommendation scores to obtain the recommendation results.

[0016] As a preferred option, it also includes: When a new micro-creative resource or a new historical propagation path is added, the propagation directed graph is updated and the graph neural network propagation is re-executed on the affected micro-creative resources, updating the graph-enhanced micro-creative resource vector, the representation in the unified semantic space, and the propagation influence prediction branch.

[0017] The second objective of this invention is to provide a blockchain-based micro-creative resource semantic understanding and intelligent recommendation system, comprising: The data acquisition module is used to acquire multimodal content data of micro-creative resources and their historical propagation paths on the blockchain, and generate an initial embedding vector. The graph propagation enhancement module is used to construct a propagation directed graph with micro-creative resources as vertices based on the historical propagation path, and to integrate the propagation behavior features into the vertex representation through a graph neural network to obtain a graph-enhanced micro-creative resource vector. A cross-modal alignment module is used to map the graph-enhanced micro-creative resource vectors to a unified semantic space through contrastive learning; The influence constraint training module is used to introduce propagation influence constraints into the unified semantic space for model training, and to add a propagation influence prediction branch to obtain the predicted influence of the micro-creative resources. The recommendation and incremental update module is used to receive queries, encode the queries and map them to the unified semantic space, perform similarity matching with candidate micro-creative resources, combine the predicted influence to obtain recommendation results, and perform incremental updates when new micro-creative resources or new historical propagation paths are added.

[0018] A third objective of this invention is to provide an electronic device comprising: A memory for storing computer programs, the memory including non-volatile memory and random access memory, the non-volatile memory for storing system programs and the random access memory for caching runtime data; A processor is used to execute programs stored in memory to implement any of the steps of a blockchain-based micro-creative resource semantic understanding and intelligent recommendation method.

[0019] The fourth objective of this invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of any of the aforementioned blockchain-based micro-creative resource semantic understanding and intelligent recommendation methods.

[0020] The beneficial effects of this invention are as follows:

[0021] 1. This invention constructs a directed graph for the propagation of micro-creative resources, integrating users' continuous attention trajectory, repeated attention counts, propagation popularity, and time decay into graph attention propagation. This allows the graph to enhance the micro-creative resource vectors, simultaneously carrying content semantics and propagation relationships. Consequently, among similar content, micro-creative resources with higher actual attention and more recent activity are prioritized for recommendation, improving the matching degree between recommendation results and user attention behavior.

[0022] 2. This invention utilizes the behavior of the same user continuously focusing on micro-creative resources of different modalities within a short time window to automatically construct cross-modal positive sample pairs. It can establish a unified semantic space for modalities such as text, images, and audio without manual annotation, significantly reducing the annotation cost of cross-modal alignment, and can adapt to newly emerging modal combinations.

[0023] 3. This invention enables the unified semantic space to have the ability to distinguish the gradient of propagation influence by using propagation influence perception triples, dynamic margin loss, and high influence compact regularization constraints. When searching, users can prioritize the discovery of micro-creative resources with higher reference value under the same topic, thus realizing the synergistic expression of semantic similarity and propagation influence.

[0024] 4. This invention integrates the prediction of dissemination influence into the recommendation score, enabling candidate micro-creative resources with high dissemination potential to obtain higher rankings; at the same time, through the influence prediction branch, it provides potential influence references for cold-start micro-creative resources lacking attention records based on content semantics and graph dissemination relationships, realizing intelligent recommendation that integrates semantic matching and influence assessment. Attached Figure Description

[0025] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0026] Figure 1 This is a flowchart illustrating a blockchain-based method for semantic understanding and intelligent recommendation of micro-creative resources, according to an embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the process of mapping graph-enhanced micro-creative resource vectors to a unified semantic space through contrastive learning in an embodiment of the present invention; Figure 3 This is a schematic diagram of a local structure of a directed graph of micro-creative resource propagation constructed based on continuous attention trajectories of attention addresses in an embodiment of the present invention; Figure 4 This is a schematic diagram of the distribution of micro-creative resources of different modalities in a unified semantic space after modal adaptive mapping in an embodiment of the present invention; Figure 5 This is a schematic diagram of the normalized propagation influence distribution of micro-creative resources in the unified semantic space in an embodiment of the present invention; Figure 6 This is a scatter plot comparing the predicted normalized propagation influence and the actual normalized propagation influence output by the propagation influence prediction branch in this embodiment of the invention. Figure 7 This is a sorted bar chart of the candidate micro-creative resource recommendation scores returned by the creative demand query in this embodiment of the invention.

[0027] The accompanying drawings have illustrated specific embodiments of the invention, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the invention in any way, but rather to illustrate the concept of the invention to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0028] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without inventive effort are within the scope of protection of the present invention. Unless otherwise defined, the technical or scientific terms used herein should have the ordinary meaning understood by those skilled in the art to which this invention pertains.

[0029] Terminology Explanation

[0030] Micro-creative resources: refer to creative content that is recorded on the blockchain and includes at least one modality of text, image, and audio. Each micro-creative resource corresponds to a unique resource identifier.

[0031] Propagation behavior characteristics: refers to information extracted from the historical propagation paths stored in the blockchain that reflects users' continuous attention behavior, including directed edge weights, normalized heat of neighboring vertices, freshness of attention time, number of repeated attentions, and propagation direction identifiers.

[0032] Softmax normalization is a mathematical function that transforms a set of real numbers into a probability distribution. Given a K-dimensional vector z, it calculates the exponent for each component and divides it by the sum of all component exponents, ensuring that the output value falls within the interval (0,1) and the sum is 1. In this invention, it is used to calculate modal gating weights and to normalize attention coefficients in graph attention propagation.

[0033] GELU (Gaussian Error Linear Unit): A neural network activation function that determines the output probabilistically based on the input value: positive values ​​are largely retained, while negative values ​​are largely suppressed. Compared to ReLU, GELU is smoother and differentiable in the negative region, improving gradient propagation and training stability. In this invention, it is used after the fully connected layers of each projection layer and the prediction head.

[0034] Example 1

[0035] This embodiment provides a blockchain-based method for semantic understanding and intelligent recommendation of micro-creative resources. Taking micro-creative resources as the processing object, this invention constructs an intelligent recommendation process that integrates blockchain propagation records and local multimodal content for resource representation, directed propagation graph enhancement, cross-modal alignment guided by attention trajectories, clustering training constrained by propagation influence gradients, and semantic matching and influence assessment.

[0036] This invention provides a blockchain-based method for semantic understanding and intelligent recommendation of micro-creative resources, such as... Figure 1 As shown, it includes the following steps:

[0037] S1, acquire multimodal content data of micro-creative resources to generate an initial embedding vector.

[0038] In this embodiment, multimodal content data refers to one or more of the following: text, images, and audio, contained in each micro-creative resource. Specifically:

[0039] 1. Text: Natural language description of creative content, including theme, style, purpose, mood, authorization information, etc.

[0040] 2. Images: Visual content related to creativity, such as hand-drawn line drawings, digital paintings, and concept designs, used to represent visual features such as composition, shape, and texture.

[0041] 3. Audio: Audio content related to creativity, such as background music, ambient sound effects, and vocal clips, used to represent acoustic characteristics such as rhythm, timbre, and energy changes.

[0042] Simultaneously, the dissemination records of the micro-creative resources are obtained from the blockchain, and each dissemination record contains at least the following fields:

[0043] 1. Blockchain Account Identifier: A unique identifier for the project to which Micro-Creative Resources belong.

[0044] 2. Work Number: A unique numerical number assigned to each micro-creative resource within the same project.

[0045] 3. User Account: The identifier of the user account receiving the follow behavior (target account).

[0046] 4. Initiating Account: The user account identifier (source account) that initiated the follow action.

[0047] 5. Timestamp: Focuses on the time (accurate to seconds) when a behavior is recorded on the blockchain.

[0048] 6. Popularity Weight: The popularity value corresponding to this attention behavior, such as the number of visits.

[0049] 7. Record Identifier: A unique index of this propagation path on the blockchain, used for querying and verification.

[0050] After acquiring the aforementioned multimodal content data, standardization processes are performed on each modality (text cleaning, image size normalization, audio resampling, and Mel-frequency conversion). Then, gated fusion is performed on the multimodal content data. This involves using a pre-trained encoder to extract initial features for each modality, and then weighting and summing these features using a gated fusion mechanism (dynamically assigning weights based on the presence of modalities) to obtain an initial embedding vector. This initial embedding vector only encodes semantic information and does not contain blockchain propagation records.

[0051] The aforementioned propagation record data is not used in the generation of the initial embedding vector, but is instead used in the subsequent step S2 to construct the propagation directed graph. This design maintains the purity of the content representation, avoids premature coupling of propagation information, and facilitates the controlled integration of propagation behavior features during the graph enhancement stage.

[0052] Finally, by standardizing the multimodal content data and dissemination records, a micro-creative resource sample table is generated, and the initial embedding vector of each micro-creative resource is obtained through multimodal coding and gating fusion.

[0053] The specific steps are as follows:

[0054] S11: Acquire and standardize the multimodal content data and historical dissemination paths of micro-creative resources, and generate a micro-creative resource sample table.

[0055] Obtain the micro-creative resource set A. For any micro-creative resource u, read its blockchain account identifier, work number, text, image, audio, and historical dissemination path stored on the blockchain. The specific steps are as follows:

[0056] 1. Generate a unique identifier: Convert the blockchain account identifier to lowercase and remove leading and trailing spaces, convert the work number to a string, and concatenate them with "_" to form "account identifier_work number".

[0057] 2. Processing dissemination paths: Sort multiple dissemination paths of the same micro-creative resource in ascending order by timestamp to form a dissemination path sequence; record the weight of the most recent dissemination popularity. (e.g., number of visits) and the time of the last visit If there is no follow record, then and The mark is empty.

[0058] 3. Text Processing: Remove invalid control characters, repeated whitespace characters, and web page tags, and retain text related to the creative theme, purpose, style, mood, and authorization instructions; if the text length exceeds the preset maximum length (such as 512 words or characters), it will be truncated; if the text does not exist, a text modality missing identifier will be recorded.

[0059] 4. Image processing: Convert to RGB format and scale to a fixed size (e.g., 224×224); if the original ratio is different from the target size, scale by the longer side and center-fill the shorter side; normalize pixel values ​​to [0,1]; if the image does not exist, record the image modality missing identifier.

[0060] 5. Audio Processing: Convert to mono, resample to 16kHz or 22.05kHz; if the duration exceeds the preset maximum duration (e.g., 10 seconds), extract the 10-second continuous segment with the highest energy; if insufficient, pad with zeros to a fixed length; convert the processed audio to a Mel spectrogram (Mel band number 64 or 128, frame shift 10~25ms); if no audio is present, record the audio mode missing identifier.

[0061] 6. Compile a sample table of micro-creative resources: Include the unique identifier, text, image, audio, missing modality identifiers, and the weight of the most recent dissemination popularity for each micro-creative resource. The most recent transmission time The on-chain propagation sequence was compiled into a micro-creative resource sample table. .

[0062] In this embodiment, the micro-creative resource "City Night Scene in a Specific Style" is used as an example. The original text content is "Draft of a looping sound effect for a city night scene in a specific style, suitable for use as a short video intro," which is cleaned and preserved intact. The image is a 2048×2048 RGB image, which is scaled and filled to become a 224×224 pixel image, showing the outline of high-rise buildings under neon lights. The audio is 15 seconds of electronic music, which is resampled to 16kHz and the 10 seconds with the highest energy are extracted and converted into a 128-band Mel spectrogram. The blockchain historical propagation path includes two attention behaviors: the first initiating account 0x1a2b..., the second user account 0x3c4d..., which are recorded after being sorted by timestamp. =2500 (times), last time viewed =March 1, 2025. The above information has been compiled into .

[0063] S12, Encode each modality in the multimodal content data to obtain the initial features of each modality.

[0064] Input the text of the micro-creative resource u into a pre-trained text encoding model (such as BERT) to obtain the initial text features. The dimension can be 768; if there is no text, fill the vector with all zeros and set the missing text flag to 1. Initial text features Used to characterize the creative themes, uses, styles, and semantic attributes of a text.

[0065] The image of the micro-creative resource u is input into a pre-trained visual encoding model (such as ViT) to obtain the initial features of the image. The dimension can be 768; if there is no image, fill the vector with all zeros and set the image missing flag to 1. Initial image features Used to characterize the shape structure, visual style, compositional relationships, and line texture in an image.

[0066] The audio Mel-spectrum of the micro-creative resource u is input into a pre-trained audio coding model (such as AST) to obtain the initial audio features. The dimension can be 768; if there is no audio, fill the vector with all zeros and set the audio missing flag to 1. Initial audio features Used to characterize rhythm, timbre, energy changes, and acoustic style in audio.

[0067] S13, construct the modal effective vector, wherein each component of the modal effective vector corresponds one-to-one with each mode.

[0068] Constructing the effective modal vector Its dimension is 3, with three components corresponding to the text modality, image modality, and audio modality, respectively. When a modality exists, the corresponding component is set to 1; otherwise, it is set to 0. For example, if the micro-creative resource u only contains text and images, and no audio, then... If the micro-creative resource u only contains audio, then The effective modal vectors are used to reduce the impact of missing modal vectors (all zero) on the results during subsequent gating fusion.

[0069] S14. Based on the modality effective vector, assign a minimum value to the gating score corresponding to the missing modality, and normalize the gating score of the existing modality to obtain the gating weight.

[0070] Initial features of the text Initial features of the image and initial audio features Each input modality projection layer is an independent input. Each modality projection layer consists of a fully connected layer, layer normalization, and a GELU activation function, mapping the input 768-dimensional features to 256-dimensional modality projection features. (Modality effective vector) Instead of participating in the calculation of the projection layer, it is used for the generation of subsequent modality gating weights: when calculating the gating score of the modality gating weights, the score corresponding to the missing modality is set to a minimum value based on the effective modality vector, so that the weight of that modality approaches 0 after Softmax normalization, thus allowing the contribution of the existing modality to be allocated by the model. Among them, the fully connected layer is used for feature dimension transformation, layer normalization is used for stable training, and the GELU activation function is used to introduce nonlinearity.

[0071] S15, the projection features of each modality are weighted and summed according to the gating weights to obtain the initial embedding vector.

[0072] Take the effective modal vector constructed in step S13 The three modal projection features are each input into a fully connected layer to obtain the corresponding gating scores. Based on... The scores corresponding to the missing modalities are set to the minimum values, and then Softmax normalization is performed on the three scores to obtain the gating weights for text, image, and audio. The projected features of the three modalities are then weighted and summed according to the gating weights to obtain the initial embedding vector. Dimension =256.

[0073] After the above processing, each micro-creative resource obtains an initial embedding vector that integrates multimodal content. This vector carries semantic information of the content, and the gating mechanism effectively handles the problem of missing modalities, providing a unified content foundation for subsequent graph enhancement.

[0074] In this embodiment, the text of the micro-creative resource "specific style city night scene" is encoded to obtain 768-dimensional features, the image is encoded to obtain 768-dimensional features, and the audio Mel-spectral mapping is encoded to obtain 768-dimensional features. Modal effective vector. =[1,1,1]. Each modal projection layer maps 768 dimensions to 256 dimensions. The three gate scores are then processed by Softmax to obtain text gate weights of 0.4, image gate weights of 0.35, and audio gate weights of 0.25. The weighted summation yields the initial 256-dimensional embedding vector. .

[0075] S2, Based on the historical propagation path of the micro-creative resource on the blockchain, construct a propagation directed graph with the micro-creative resource as the vertex.

[0076] Multimodal coding itself cannot express the propagation relationship implied by users' continuous attention behavior. In order to integrate the behavioral information in the blockchain record into the micro-creative resource representation, this embodiment constructs a directed propagation graph, specifically: taking the micro-creative resource as the vertex, sorting the historical propagation path in chronological order according to user account, establishing directed edges between adjacent vertices and accumulating the number of repeated attentions, calculating the edge weight based on the number of repeated attentions, the propagation popularity weight in the historical propagation path, and the attention time interval in the historical propagation path, adding self-loop edges, and retaining adjacent vertices whose weights meet preset conditions.

[0077] Using each micro-creative resource in the micro-creative resource set A obtained in step S1 as a vertex, construct a directed graph Gra=(A,Er) for micro-creative resource propagation, where Er represents the set of directed edges, which can be constructed through the following process:

[0078] 1. Initialization: Er is initially an empty set.

[0079] 2. Generate adjacent directed edges: Group and chronologically sort the historical propagation paths according to user accounts, establish directed edges between adjacent micro-creative resources, and accumulate the number of repeated follows from one micro-creative resource to another. Specifically, for each pair of adjacent micro-creative resources (e.g., i and j) in each follow path, calculate the follow time interval. If you are concerned about time intervals If the propagation window does not exceed a preset period (e.g., 7 days, 30 days, or 90 days), a directed edge is established between vertex i and vertex j of the micro-creative resource, and this directed edge is added to Er. If the same pair of micro-creative resources (i→j) is generated by the trajectories of multiple user accounts, the number of repeated follows is accumulated. However, only one directed edge record is kept in Er (it is not added repeatedly).

[0080] 3. For each directed edge, calculate the edge weight based on the number of repeated follows, the popularity weight of the micro-creative resource pointed to by the directed edge, and the time decay corresponding to the time interval of continuous follows. Specifically: for each directed edge i→j, calculate the popularity compression value of micro-creative resource j. ,in Let `j` be the most recent dissemination popularity weight (e.g., number of visits). Simultaneously, calculate the maximum popularity compression value for all micro-creative resources with dissemination popularity weights across the entire graph. (If none of the micro-creative resources have a popularity record, then) Let's set it to 1). Then calculate the edge weight of the directed edge i→j. : ; Where: log represents the common logarithm to base 10, The preset time decay coefficient is used to control the rate at which the influence of historical attention behavior decreases as the time interval increases, with a value ranging from 0.01 to 0.05. This is a term for compressing the number of times a user repeatedly follows an item, which is used to make the edge weight stronger when multiple users repeatedly follow the item, while avoiding the infinite amplification of high-frequency edges. The "time freshness" factor indicates that the shorter the interval between attention, the stronger the association. This represents the normalized heat level of micro-creative resource j.

[0081] For example, if the directed edge i→j is formed by 4 user accounts, then =4; The most recent dissemination popularity weight of micro-creative resource j is 2500 (times), and the popularity compression value is 4. =ln(2501)≈7.82; Maximum heat compression value of the entire map =9.21; Time interval for monitoring =3 days; Time decay coefficient =0.02; then the edge weight =log(5)×(7.82 / 9.21)×exp(-0.02×3)≈0.699×0.849×0.942≈0.559. If the attention interval becomes 60 days, then the time freshness term exp(-0.02×60)=exp(-1.2)≈0.301, and the edge weight drops to about 0.178, indicating that the association of continuous attention behavior over a long time interval is significantly weakened.

[0082] It's important to note that the directed graph for the dissemination of micro-creative resources is not built based on content similarity, but rather on a combination of factors including continuous following behavior of user accounts, the weight of dissemination popularity, the time interval between following actions, and the number of times repeated following occurs. Based on this, information from blockchain records such as "who drives whom," "which micro-creative resources appear under the same attention intent," and "which recently popular micro-creative resources have dissemination influence" can be transformed into a computable graph structure.

[0083] 4. Add self-loop edges: Add self-loop edges (vertices pointing to themselves) to each of the micro-creative resources, and also add Er, with the edge weight set to 1.

[0084] 5. Retain adjacent micro-creative resources whose weights meet preset conditions: For each micro-creative resource, retain a preset number of outgoing and incoming neighbors with the highest weights. To reduce computational load, retain the top-weighted neighbors for each micro-creative resource. A neighbor and the former One of the neighbors on the side, Examples of values ​​are 20 or 50. This processing is used to limit overly dense connections from highly active user accounts, reduce training computation, and minimize the noise impact of weakly associated edges on graph propagation.

[0085] In one embodiment, such as Figure 3 As shown, this illustrates the local structure of the directed graph of micro-creative resource dissemination. Figure 3 The graph is schematically drawn using micro-creative resources and directed edges. Vertices A1 to A8 represent micro-creative resources, and directed edges (arrows) represent the direction of a user account's continuous attention trajectory. Edge weights and attention time intervals are labeled on the edges, and examples of self-loop edges are included. This graph is not built based on content similarity, but rather on the propagation behavior stored on the blockchain, which can reflect the "who drives whom" propagation relationship and provide a structural foundation for subsequent graph attention propagation.

[0086] In this embodiment, we take the micro-creative resource "City Night Scene in a Specific Style" (denoted as Micro-creative Resource A) as an example. Its blockchain historical propagation path shows that user account 0x1a2b... followed A on March 1, 2025, followed Micro-creative Resource B (another audio in a specific style) on March 3, and followed Micro-creative Resource C (accompanying video material) on March 5. Therefore, a directed edge A→B is established between A and B (following interval of 2 days), and a directed edge A→C is established between A and C (following interval of 4 days). The most recent propagation popularity weight for Micro-creative Resource B is 3000 times, and for C it is 1200 times. The maximum popularity compression value for the entire graph is... Taking 5000 instances from a highly popular micro-creative resource, we calculated S(3000) = ln(3001) ≈ 8.01 and S(1200) = ln(1201) ≈ 7.09. =ln(5001)≈8.52. Edge weight calculation: = log(1+1)·(8.01 / 8.52)·exp(-0.02×2) ≈0.301×0.940×0.961 ≈ 0.272; = log(1+1)·(7.09 / 8.52)·exp(-0.02×4) ≈0.301×0.832×0.923 ≈ 0.231. Furthermore, a self-loop edge (weight 1) is added to each vertex. The final result is a propagating directed graph.

[0087] S3, using the initial embedding vector as the initial feature, the propagation behavior feature is integrated into the vertex representation through a graph neural network to obtain a graph-enhanced micro-creative resource vector.

[0088] On the constructed directed graph of propagation, a feature vector containing the edge weights and the directions of the directed edges is constructed. This feature vector is mapped to a behavior bias and superimposed on the content relevance score. Normalization yields an attention coefficient. The features of neighboring vertices are then weighted and aggregated according to this coefficient to update the features of the center vertex, resulting in a graph-enhanced micro-creative resource vector. The steps include:

[0089] S31, Construct the feature vector of propagation behavior.

[0090] This step constructs a propagation behavior feature vector to prepare for subsequent bias stacking, normalization, and weighted aggregation. The propagation behavior features include: the weights of directed edges in the propagation directed graph, the normalized popularity of the adjacent micro-creative resources pointed to by the directed edges, the temporal freshness of the attention time corresponding to the directed edges, the compression term of the repeated attention count of the directed edges, and the propagation direction identifier of the directed edges.

[0091] The initial embedding vector obtained in step S15 This serves as the input to the first layer of the graph attention network. Let the input vertex features of the l-th layer (l is counted from 1) be... The output vertex features are For each propagation neighbor vertex v (including self-loops) of the central vertex u, construct a propagation behavior feature vector. The vector contains 5 components:

[0092] The first component is the directed edge weight: if there exists an edge from u to v, then take... If there exists an edge from v to u, then take it. When the loop is in place, take 1. It describes the intensity of the behavior that propagates from u to v (u affects v). It describes the intensity of the behavior from v to u (v influences u). This weight comprehensively reflects the number of times repeated attention is received, the popularity level of the target micro-creative resource, and the freshness of time. The larger the value, the stronger the propagation association between u and v.

[0093] When calculating the attention coefficient, the central vertex u needs to consider both its neighbors (incoming edges) and its neighbors (outgoing edges) because propagation can be bidirectional. For example, a micro-creative resource u may be influenced by a micro-creative resource v with high propagation influence (incoming edge), and u may also influence subsequent micro-creative resources (outgoing edge). Therefore, the correct edge weights must be selected based on the direction; otherwise, the propagation direction will be confused, leading to incorrect relationships learned by the graph attention network. In this embodiment, the outgoing neighbors are... For neighboring countries within the border Then, along with the direction identifier, it is input into the behavior bias mapping layer, allowing the model to learn the contribution of different directions.

[0094] The second component is the normalized heat of neighboring micro-creative resources: that is, the normalized heat level of the neighboring vertex v, as shown in the formula. ,in This is the thermal compression value. This is the maximum heat compression value for the entire graph. The value ranges from (0,1) and represents the historical dissemination heat of neighboring micro-creative resources v. The higher the dissemination heat, the closer this value is to 1, indicating that v has a higher dissemination influence; the lower the dissemination heat, the closer this value is to 0. The purpose of introducing this component is to allow micro-creative resources with high dissemination influence to obtain higher attention weight in dissemination, reflecting the heat phenomenon that "high-heat micro-creative resources are more influential."

[0095] The third component is time freshness: the time freshness corresponding to the time of interest, calculated using the following formula: ,in This is the time decay coefficient, with a value range of [0.01, 0.05]. The interval (in days) for continuous attention is defined. This component ranges from (0,1). The shorter the attention interval, the closer the freshness is to 1, indicating that recent attention behavior contributes more to the current association. The longer the interval, the freshness index decays, indicating that the association of attention behavior from long ago is weakened. This design enables the model to pay more attention to recent hot topics and trend changes, avoiding interference from outdated attention behavior.

[0096] The fourth component is the repeated attention count compression term: the formula is... ,in This component represents the number of different user accounts that repeatedly perform the continuous follow from u to v. The value ranges from (0, +∞), but logarithmic compression slows down the growth of the number of repetitions. For example, 1 repetition corresponds to log(2) ≈ 0.301, and 4 repetitions correspond to log(5) ≈ 0.699. This design reflects the strength of "group consensus": the more independent user accounts switch from u to v, the more recognized the connection between u and v is, and the higher the edge contribution should be assigned; at the same time, it avoids the infinite amplification of weights caused by a few high-frequency abnormal addresses.

[0097] The fifth component is the propagation direction identifier: 1 for outgoing edges (u→v), -1 for incoming edges (v→u), and 0 for self-loops. This component is used to explicitly distinguish the propagation direction, enabling the model to learn the different effects of forward propagation (from u to v) and backward propagation (from v to u). For example, propagation from micro-creative resources with high propagation influence to micro-creative resources with low propagation influence may represent a "leading" relationship, while backward propagation may represent a "following" relationship. The direction identifier helps the behavior bias mapping layer distinguish between these two cases.

[0098] S32, Input the propagation behavior feature vector into the learnable behavior bias mapping layer to obtain the behavior bias scalar.

[0099] Will Input behavior bias mapping layer to obtain behavior bias scalar The behavior bias mapping layer can be implemented in two ways: one is by a single fully connected layer with an input dimension of 5 and an output dimension of 1; the other is by a two-layer fully connected network, where the first layer outputs a dimension of 16 followed by a GELU activation function, and the second layer outputs a single scalar. Using a learnable behavior bias mapping layer allows the model to learn the influence of behavioral factors such as "recent popularity," "repeated attention by multiple users," and "positive propagation" on attention scores, rather than relying entirely on manually fixed weights.

[0100] S33, calculate the content relevance score between the center vertex and its neighboring vertices, superimpose the behavior bias scalar, and obtain the attention coefficient by processing with a normalized exponential function.

[0101] Perform linear transformations on the central vertex u and its neighboring vertex v respectively: and Multiply by the learnable feature transformation matrix of the l-th layer The transformation features are obtained. Used to transfer input vertex features from Dimension mapping to dimension; This represents the dimension of the input features at layer l; This represents the dimension of the output features of layer l, with examples of values ​​of 256 or 128. Then, the unnormalized attention score is calculated: the center vertex transformation feature is concatenated with the neighboring vertex transformation features, and the content relevance score is obtained through the learnable attention vector, plus the behavior bias scalar. Finally, it is activated by LeakyReLU.

[0102] Perform Softmax normalization on all neighboring vertices (including self-loop vertices) of the central vertex u to obtain the attention coefficients. This ensures that the sum of all attention coefficients corresponding to the same central vertex is 1.

[0103] S34, The attention coefficient is used to perform weighted aggregation of the features of neighboring vertices, update the features of the central vertex, and obtain the graph-enhanced micro-creative resource vector through multi-layer propagation.

[0104] Attention coefficients are used to weight and aggregate the transformation features of neighboring vertices to obtain weighted aggregated neighbor features. These weighted aggregated neighbor features are then added to the self-loop transformation features of the center vertex, and after passing through an ELU activation function and layer normalization (optional residual connections), the vertex features output by the l-th layer are obtained. Stacking layer( (Example value is 2), ultimately resulting in a graph-enhanced micro-creative resource vector. = Its dimensions Examples of possible values ​​are 128 or 256.

[0105] For example, suppose there are two propagating neighbor vertices of the central vertex u. and . The corresponding micro-creative resources had high dissemination popularity (e.g., 5000 views), and the attention time was only one day apart from u, while multiple user accounts showed activity from u to u. Continuous attention behavior; The propagation frequency is low (e.g., 200 times) and the propagation interval is 120 days. Under these circumstances, the propagation behavior feature vector... The heat component, time freshness component, and repeated attention component are all high, and the behavioral bias scalar obtained after the behavioral bias mapping layer is also high. Larger, thus improving Attention score; and Medium heat content, low time freshness, behavioral bias scalar Smaller, even The content of u is highly similar. It will also result in a lower attention score.

[0106] In this embodiment, following the propagation directed graph constructed in step S2, micro-creative resource A is taken as the central vertex, and its propagation neighbors are B and C (self-loop edges are considered as their own neighbors). In step S31, propagation behavior feature vectors are constructed for A and B, and A and C respectively. and This includes: directed edge weights ( =0.272, =0.231); neighboring vertices Values ​​(B is 8.01 / 8.52≈0.94, C is 7.09 / 8.52≈0.83); Time freshness (exp(-0.02×2)≈0.96, exp(-0.02×4)≈0.92); Repeat attention count compression term (both are log2≈0.301); Direction identifier (outgoing edge is 1). After the behavior bias mapping layer, because B has higher popularity and shorter time interval, its behavior bias scalar is greater than C's behavior bias scalar. In step S33, the content attention score is calculated and the bias is superimposed. After normalization, the attention coefficients are obtained: A's attention coefficient to B is about 0.6, A's attention coefficient to C is about 0.3, and A's attention coefficient to itself is about 0.1. In step S34, the weighted aggregation of neighbor features is passed through 2 layers of propagation to obtain the graph enhancement vector of micro-creative resource A. With a dimension of 256, this vector incorporates information about B's high dissemination influence.

[0107] S4, through contrastive learning, the graph-enhanced micro-creative resource vectors are mapped to a unified semantic space.

[0108] In existing technologies, micro-creative resources of different modalities differ significantly in their original content space: text leans towards conceptual expression, images towards structural expression, and audio towards acoustic expression. Directly comparing graph-augmented micro-creative resource vectors of different modalities easily leads to the problem that text micro-creative resources are only similar to text micro-creative resources, image micro-creative resources are only similar to image micro-creative resources, and audio micro-creative resources are only similar to audio micro-creative resources, making cross-modal recommendation difficult. To solve this problem, cross-modal positive sample pairs are constructed using the historical propagation paths of the same user account across different micro-creative resources. These cross-modal positive sample pairs are used as positive samples for contrastive learning, bringing the two graph-augmented micro-creative resource vectors corresponding to the cross-modal positive sample pairs closer together in the unified semantic space. Specifically, this is achieved through the following sub-steps, such as... Figure 2 As shown:

[0109] S41, Input the graph-enhanced micro-creative resource vector into the shared projection layer to obtain the shared basic vector.

[0110] The graph-enhanced micro-creative resource vector obtained in step S3 Input a shared projection layer to obtain shared fundamental vectors. , dimension (Example: 128). The shared projection layer consists of a fully connected layer, layer normalization, and a GELU activation function. Vitagram enhances micro-creative resource vector mapping to A shared fundamental vector is used. The shared projection layer is used to preserve the general semantic information in the graph-enhanced micro-creative resource vector and the propagation context information stored in the blockchain.

[0111] S42, select the corresponding modality-specific projection module based on the main modality of the micro-creative resource, and output the modality compensation vector.

[0112] Based on the primary modality of the micro-creative resource u, a corresponding modality-specific projection (MSP) module is selected. The MSP module is used to learn offset compensations for different modalities in a unified semantic space. Text modality, image modality, audio modality, and different combinations of modalities can each be configured with an independent MSP module. Each MSP module employs a two-layer fully connected network. The output dimension of the first layer is taken as or 2 Followed by the GELU activation function; Layer 2 output 3D modal compensation vector Different primary modalities correspond to different parameters, allowing text, images, and audio to learn their own distribution correction methods separately.

[0113] Shared fundamental vectors Input the corresponding modality-specific projection module to obtain the modality compensation vector. It is used to correct embedding bias caused by differences in the primary modes.

[0114] S43, add the shared fundamental vector and the modal compensation vector, and perform layer normalization on the addition result to obtain the representation in the unified semantic space.

[0115] Then, the shared fundamental vector is added to the modal compensation vector, and layer normalization is performed to obtain the representation in the unified semantic space:

[0116] in, The representation layer normalization operation is used to stabilize the numerical distribution of representations in a unified semantic space. This residual structure can perform lightweight compensation for different modalities while preserving the backbone information of shared basic vectors, avoiding the complete coverage of the propagation network in the graph-enhanced micro-creative resource vectors by modality-specific projection modules.

[0117] S44. Determine the main modality of each micro-creative resource based on the modality effective vector, and find micro-creative resource pairs that are followed sequentially within a preset time window and have different main modalities, as cross-modality positive sample pairs.

[0118] Based on the effective modal vector obtained in step S13 The dominant modality of each micro-creative resource u is determined. The dominant modality represents the content type that predominates in terms of information content within the micro-creative resource, and can include text modality, image modality, audio modality, text-image bimodality, text-audio bimodality, image-audio bimodality, and text-image-audio trimodality. Specifically, if the micro-creative resource contains only text, the dominant modality is text modality; if it contains only images, the dominant modality is image modality; if it contains only audio, the dominant modality is audio modality; and if it contains both text and images, the dominant modality is text-image bimodality. If further differentiation of information content is needed, the dominant modality can be determined based on text length, image clarity, and audio energy length.

[0119] Based on user account following history, search for micro-creative resources (u and v) that have been followed sequentially within a preset cross-modal time window (e.g., 48 hours, 7 days, or 14 days) and have different primary modalities. If such continuous following behavior exists, then... Add cross-modal positive sample micro-creative resource pairs It should be noted that the shorter the time window, the more reliable the cross-modal positive micro-creative resource pairs are, but the number may be smaller; the longer the time window, the more positive micro-creative resource pairs are, but the weak correlation noise will also increase.

[0120] right Deduplication is performed: If the same user account creates the same micro-creative resource pair multiple times, or multiple user accounts create the same micro-creative resource pair, only one positive sample micro-creative resource pair is retained, and the frequency of occurrence of this positive sample micro-creative resource pair is recorded. This frequency can be used as a sample weight during training (the more frequent the occurrence, the more stable the cross-modal association). Abnormal cross-modal positive sample micro-creative resource pairs are also filtered: If the attention interval between two micro-creative resources exceeds a preset cross-modal time window, or if either micro-creative resource is marked as an invalid micro-creative resource, a removed micro-creative resource, or an abnormal record, it will not be added. .

[0121] It should be noted that the cross-modal positive sample micro-creative resource set Instead of being manually labeled, these annotations are generated from user attention behavior stored on the blockchain. Based on this, decentralized real-world attention trajectories can be used to establish weakly supervised alignment relationships between text, images, and audio, reducing the cost of manual annotation.

[0122] In one example of this embodiment, if user account b follows a "forest healing style short copy" on day 1 and a "forest environment white noise audio" on day 2, and the main modalities of the two micro-creative resources are text modality and audio modality respectively, then the two micro-creative resources constitute a cross-modal positive sample micro-creative resource pair, indicating that the same user account has continuously followed micro-creative resources of different modalities but possibly the same theme in a short period of time.

[0123] S45, using the cross-modal positive sample pair as positive samples for contrastive learning, the two representations corresponding to the cross-modal positive sample pair are brought closer in the unified semantic space.

[0124] For each cross-modal positive sample micro-creative resource pair ,calculate and Cosine similarity between them. Selecting cross-modal negative sample micro-creative resources from the same training batch: for positive sample micro-creative resources... For micro-creative resource u, other micro-creative resources that differ from its primary modality and do not form a cross-modal positive sample micro-creative resource pair with u are selected as cross-modal negative sample micro-creative resources, denoted as . A cross-modal contrastive training method is employed to optimize the unified semantic space, resulting in closer proximity between positive cross-modal micro-creative resource pairs and greater distance between negative cross-modal micro-creative resource pairs. The cross-modal alignment loss is denoted as... In the specific calculation, temperature scaling (temperature coefficient) is applied to the cosine similarity between representations in the unified semantic space. (Examples of values ​​are 0.05 to 0.2), and comparative learning is performed using in-batch negative sample micro-creative resources. The smaller the temperature coefficient, the more the model emphasizes the discriminative power between positive and negative sample micro-creative resources.

[0125] It should be noted that this step utilizes cross-modal weak supervision signals formed by attention trajectories stored on the blockchain, rather than relying on manually established one-to-one text-image, text-audio, or image-audio pairings. By sharing a projection layer and modality-specific projection modules, the representation in the semantic space is unified while preserving content semantics, propagation context, and cross-modal alignment relationships, providing a unified foundation for subsequent propagation influence perception recommendations.

[0126] In one embodiment, a two-dimensional semantic space visualization is obtained through PCA projection. For example... Figure 4 As shown, Figure 4 Coloring by primary modality (text modality, image modality, audio modality, etc.), different colors and markers distinguish modality types, demonstrating that after modality-specific projection, micro-creative resources of different modalities are mixed and distributed in a unified space, rather than being strictly separated by modality, indicating that the model can achieve cross-modal semantic alignment. For example... Figure 5 As shown, Figure 5 Coloring by normalized dissemination influence (green to red indicates from low to high dissemination influence) shows that micro-creative resources with high dissemination influence tend to form clustered areas, confirming that the unified semantic space has the ability to express the gradient of dissemination influence through subsequent dissemination influence-sensitive clustering loss.

[0127] In this embodiment, continuing the previous example, the same user account 0x1a2b... followed text-modal micro-creative resource A (specific style text) on March 1, 2025, followed audio-modal micro-creative resource B (specific style audio) on March 2, and followed image-modal micro-creative resource C (specific style image) on March 3. The primary modalities of micro-creative resources A, B, and C are text, audio, and image modal, respectively, and they are all different from each other. Furthermore, they were followed consecutively within a 48-hour window, thus forming positive sample micro-creative resource pairs (A,B), (A,C), and (B,C), all of which were added to... During training, the graph augmentation vectors of micro-creative resource A... Input shared projection layer to obtain (128-dimensional), since its primary modality is text, a text-specific projection module is selected to obtain... According to the formula We obtain the representation in the unified semantic space. Similarly, we obtain... and .calculate and The cosine similarity is used, and audio micro-creative resources that are different from the main modality of A and do not form positive sample micro-creative resource pairs are selected from the batch as negative sample micro-creative resources (e.g., ambient sounds of other themes). Cross-modal contrastive loss. make and The similarity with negative sample micro-creative resources increased, while the similarity with negative sample micro-creative resources decreased. After multiple rounds of training, the results obtained by querying "specific style" using text were... Able to , It has a high cosine similarity, thus matching the audio micro-creative resource B and the image micro-creative resource C.

[0128] S5, introduce the propagation influence constraint into the unified semantic space for model training, and add a propagation influence prediction branch to obtain the predicted influence of the micro-creative resource.

[0129] Micro-creative resource recommendations not only need to determine the similarity of creative content, but also need to provide a reference for dissemination influence. Training only a semantic similarity model might group micro-creative resources with similar content but significantly different dissemination influence together; sorting only by historical popularity would ignore the creative demand itself. To solve these problems, this invention introduces normalized dissemination influence into a unified semantic space, constructs dissemination influence perception hierarchical triples, and trains a dissemination influence prediction head, enabling the representation in the unified semantic space to simultaneously possess cross-modal semantic comparability and the ability to express dissemination influence gradients. The specific steps are as follows:

[0130] S51 divides micro-creative resources into high, medium, and low-impact sets based on their dissemination popularity weight.

[0131] Weights of the spread popularity in the historical spread paths. Micro-creative resources u, computational normalization of dissemination influence First, calculate the thermal compression value. Then use the maximum heat compression value of the entire image. Normalization is performed to obtain The value range is [0,1]. A higher value indicates a higher historical dissemination influence of the micro-creative resource u. If the micro-creative resource u has no dissemination popularity record, then... If marked as null, this micro-creative resource will not participate in the regression training for dissemination influence, but it can still participate in cross-modal alignment training and recommendation candidate ranking.

[0132] Based on normalized dissemination influence, micro-creative resources are divided into three sets: high dissemination influence, medium dissemination influence, and low dissemination influence. Specifically, they can be categorized according to the quantile of normalized dissemination influence: micro-creative resources with the highest normalized dissemination influence are assigned to the high dissemination influence set. The middle 50% are classified into the medium-influence set, and the bottom 30% are classified into the low-influence set; alternatively, a fixed threshold can be used for classification, for example... Values ​​≥0.7 are classified into the high dissemination influence set, and values ​​≤0.3 are classified into the high dissemination influence set. When the value is less than 0.7, it is included in the medium-level communication influence set. When the value is less than 0.3, it is classified into the low propagation influence set.

[0133] S52: Select anchor points from the high or medium influence set, select positive samples from resources with the same or higher influence level as the anchor points, and select negative samples from resources with lower influence levels than the anchor points to form triples.

[0134] Anchor micro-creative resources 'a' are selected from sets of high or medium dissemination influence with existing dissemination popularity records. Priority is given to micro-creative resources with existing dissemination popularity records and complete dissemination paths stored on the blockchain. Positive sample micro-creative resources 'b' are selected for anchor micro-creative resource 'a': First, the cosine similarity between 'a' and other micro-creative resources is calculated based on their representations in the unified semantic space. The resources with the highest similarity are selected. One candidate micro-creative resource ( (Example values ​​are 20 to 50). Then, from these candidate micro-creative resources, micro-creative resources with the same or higher influence level as 'a' are selected. Finally, it is further required that the time interval between the most recent attention of 'a' and the most recent attention of 'a' does not exceed a preset window of the same period (e.g., 30 days or 90 days) to ensure that positive sample micro-creative resources reflect similar dissemination cycles. Negative sample micro-creative resources 'e' are selected for anchor micro-creative resource 'a': negative sample micro-creative resources are selected from micro-creative resources with an influence level lower than 'a', prioritizing micro-creative resources with high semantic similarity to 'a' but lower influence level as difficult negative sample micro-creative resources. Anchor micro-creative resource 'a', positive sample micro-creative resource 'b', and negative sample micro-creative resource 'e' form a dissemination influence perception level triplet (a, b, e), and all triplets are added to the dissemination influence perception level triplet set. .

[0135] In one embodiment, if anchor micro-creative resource 'a' is a high-impact, style-specific image micro-creative resource, positive sample micro-creative resource 'b' is a high-impact, style-specific short video intro audio micro-creative resource, and negative sample micro-creative resource 'e' is a common image micro-creative resource with the same style tag but lower popularity, then (a, b, e) constitutes a perceptual triplet for dissemination influence. During training, the model will push anchor micro-creative resource 'a' closer to positive sample micro-creative resource 'b' and push anchor micro-creative resource 'a' to maintain a greater distance from negative sample resource 'e'.

[0136] It should be noted that the perception triplet for dissemination influence is not a random triplet, but rather considers representational similarity, normalized dissemination influence, and proximity of attention time in a unified semantic space. Based on this, the dissemination logic of "aggregating micro-creative resources with the same theme, contemporaneous time, and high dissemination influence; and distinguishing micro-creative resources with the same theme but low dissemination influence" can be injected into the unified semantic space.

[0137] S53, adjust the contrast loss margin according to the influence difference of the triples, and apply a compact regularization constraint to the representation in the high influence set, wherein the margin includes the product of the base margin and the dynamic adjustment term.

[0138] This step is used to train a unified semantic space through influence-sensitive dynamic margins and to provide potential influence references for cold-start micro-creative resources by propagating an influence prediction head. The specific steps are as follows: For any triple (a, b, e) at the perception level of dissemination influence, calculate the embedding distance between the anchor micro-creative resource a and the positive sample micro-creative resource b. And the embedding distance between anchor micro-creative resource a and negative sample micro-creative resource e. The embedding distance can be calculated using either 1 minus cosine similarity or Euclidean distance squared. In this embodiment, 1 minus cosine similarity is used as the embedding distance.

[0139] The influence-sensitive dynamic margin is calculated based on the normalized propagation influence difference within the triplet. When the normalized propagation influence of anchor micro-creative resource a and positive sample micro-creative resource b is significantly higher than that of negative sample micro-creative resource e, the margin should be larger; when the influence difference is small, the margin should be smaller. Specifically, the propagation influence margin adjustment term is first calculated. :

[0140] in, This represents the normalized dissemination influence of anchor micro-creative resource 'a'; This represents the normalized dissemination influence of positive sample micro-creative resource b; This indicates the operation of retrieving the maximum value; This represents the normalized dissemination influence of negative sample micro-creative resource e; ε represents the moderating strength coefficient for differences in dissemination influence, with an example value of 0.3; ε represents a very small positive number to prevent the denominator from being zero, with an example value of 10. -6 By taking the maximum value, the margin will not be erroneously amplified when the dissemination influence of negative sample micro-creative resources is not lower than the average level of anchor point micro-creative resources and positive sample micro-creative resources.

[0141] Then calculate the triplet contrast loss, denoted as . The anchor point micro-creative resource a is used to constrain the anchor point micro-creative resource a to be closer to the positive sample micro-creative resource b, and the anchor point micro-creative resource a to be farther away from the negative sample micro-creative resource e.

[0142] In one implementation, the triple contrastive loss can be expressed as:

[0143] Where μ represents the basic margin hyperparameter, used to define the basic requirement for the distance difference between positive and negative sample micro-creative resources, with values ​​ranging from 0.3 to 0.8. If the distance between anchor micro-creative resource a and negative sample micro-creative resource e is already greater than the distance between anchor micro-creative resource a and positive sample micro-creative resource b by a sufficient margin (i.e., ... If the sum is less than or equal to 0, then the contrast loss of the triple is 0; otherwise, the model continues to adjust the representation in the unified semantic space through backpropagation.

[0144] Apply a compact regularization constraint to the set of micro-creative resources Ha with high dissemination impact. Calculate the cluster center of micro-creative resources with high dissemination impact. This refers to the dimension-wise average vector of the representations of high-impact micro-creative resources in the unified semantic space of the current training batch, used to characterize the central position of high-impact micro-creative resources in the unified semantic space. Then, the representation of each high-impact micro-creative resource in the unified semantic space and the center of the high-impact micro-creative resource cluster are calculated. The distances between them are calculated, and the average of these distances is used as a compact regularization term for high dissemination influence. This regularization term is used to reduce the risk of excessive dispersion of high dissemination influence micro-creative resources in a unified semantic space, enabling high dissemination influence micro-creative resources to form a more stable cluster of dissemination influence.

[0145] Combining all triple contrast loss terms with the high propagation influence compact regularization term yields the propagation influence clustering loss. Specifically, the set of triplets representing the perceived levels of dissemination influence. The average of the triplet contrast loss is then added, along with a high propagation influence compact regularization term multiplied by a compactness coefficient. ( The value range is from 0.05 to 0.2. If... Excessive size may lead to over-compression of thematic differences among highly influential micro-creative resources; if If the size is too small, the micro-creative resource clusters with high dissemination influence will lack sufficient constraints.

[0146] S54 predicts influence by propagating the influence prediction branch and outputting the prediction influence based on the representation in the unified semantic space.

[0147] After obtaining the representation in the unified semantic space, the propagation influence prediction head is connected. It is used to predict the normalized dissemination influence of micro-creative resources based on representations in a unified semantic space, and to provide a reference for the potential influence of cold-start micro-creative resources without historical dissemination popularity records.

[0148] In one implementation, the propagation influence prediction head uses a two-layer fully connected network: The output dimension of the first layer is either 64 or 128, followed by the GELU activation function; The second layer outputs a scalar, connected to a Sigmoid function, ensuring the predicted propagation influence is within the range of 0 to 1. The predicted propagation influence of the micro-creative resource u is denoted as... .

[0149] The dissemination influence prediction head is trained using micro-creative resources with historical popularity. The set of micro-creative resources with historical dissemination popularity is denoted as... For each belonging to Micro-creative resources u, calculating and predicting dissemination influence With the influence of normalized communication The squared error between the two values ​​is then averaged over all micro-creative resources with historical dissemination popularity to obtain the regression loss of dissemination influence. .

[0150] Cross-modal alignment loss Clustering loss of dissemination influence and the loss of influence in communication Combining the results, we obtain the total loss of the joint optimization, denoted as... , This represents the overall goal of the entire model training phase, used to simultaneously optimize cross-modal alignment, propagation influence gradient clustering, and propagation influence prediction capabilities.

[0151] The total loss of joint optimization can be expressed as:

[0152] Here, ξ represents the trade-off coefficient for the regression loss of dissemination influence, used to control the contribution of the dissemination influence prediction objective to the total loss of joint optimization, with values ​​ranging from 0.3 to 0.8. If a greater emphasis is needed on the reference capability of dissemination influence, ξ can be increased appropriately; if a greater emphasis is needed on cross-modal semantic alignment, ξ can be decreased appropriately.

[0153] The model is trained using an adaptive moment estimation optimizer, with an initial learning rate of 10. -4 Up to 10 -3 The batch size can be 64 or 128, and the number of training rounds can be 50 to 200. During training, each batch simultaneously samples cross-modal positive sample micro-creative resource pairs and propagation influence perception level triples, so that the unified semantic space simultaneously receives cross-modal constraints and propagation influence gradient constraints.

[0154] In one embodiment, such as Figure 6 As shown, the scatter plot compares the predicted normalized dissemination influence and the actual normalized dissemination influence output by the dissemination influence prediction head. The horizontal axis represents "actual normalized dissemination influence" (0–1), and the vertical axis represents "predicted normalized dissemination influence" (0–1). The black dashed line represents the ideal prediction line. The scatter points in the plot are roughly distributed along the diagonal, indicating a strong correlation between the predicted and actual values. This demonstrates that the dissemination influence prediction head can learn dissemination influence information related to historical attention from representations in a unified semantic space, showcasing the feasibility of providing potential dissemination influence references for cold-start micro-creative resources.

[0155] In this embodiment, we continue to use micro-creative resources A (2500 times of dissemination popularity weight, normalized dissemination influence 0.65), B (3000 times, normalized dissemination influence 0.71), and D (800 times, normalized dissemination influence 0.35) as examples. Micro-creative resource A is used as anchor micro-creative resource a, B as positive sample micro-creative resource b (higher influence and attention time all in early March), and D as negative sample micro-creative resource e. Calculation... + -2 =0.65+0.71-0.70=0.66, ω=1+0.3×0.66 / (0.65+0.71)≈1+0.198 / 1.36≈1.146.

[0156] Taking the base margin μ = 0.5, the dynamic margin is 0.573. Triple contrast loss.

[0157] During training, Being pulled closer, The distance is pushed back by at least 0.573. Simultaneously, a compact regularization is applied to the set of high-impact micro-creative resources Ha (including micro-creative resource B and other high-impact micro-creative resources), causing them to shrink towards the cluster center. The impact prediction head is based on the micro-creative resource A... Output The value is approximately 0.63, close to the actual value of 0.65. For micro-creative resources without a recorded spread popularity (such as newly released micro-creative resources with a specific style), the spread influence prediction head can output the predicted spread influence as a cold start reference. The total loss is calculated through joint optimization. middle, From step S45, The sum of the triple contrast loss and the high propagation influence compact regularization term from this step. The mean squared error from the propagation influence prediction head is set to 0.5. After multiple rounds of training, high-propagation-influence micro-creative resources (such as micro-creative resource B) and low-propagation-influence micro-creative resources (such as micro-creative resource D) in the unified semantic space form a propagation influence gradient distinction under the premise of semantic similarity.

[0158] S6. Receive a query, encode the query and map it to the unified semantic space to obtain a query representation, select candidate resources from the micro-creative resources, perform similarity matching between the query representation and the representation of the candidate resources, and combine the predicted influence to obtain a recommendation result.

[0159] This embodiment receives creative requirement queries input by users, and the query format includes one or more of text, images, and audio. The query is encoded and mapped to a unified semantic space to obtain its representation in the unified semantic space. The similarity between the query and the representations of each candidate micro-creative resource in the unified semantic space is calculated. A recommendation score is calculated based on the predicted influence of each candidate micro-creative resource, and a recommendation list is output after sorting by score. The specific steps are as follows:

[0160] S61, calculate the cosine similarity between the query representation and the candidate resource representation.

[0161] The system retrieves creative request queries (q) entered by users, which can include one or more of the following: text, images, and audio. For example, a user could enter "low ambient sound suitable for the opening of a sci-fi short film," or upload an image with a brief text description.

[0162] The query q is processed using the same procedure as steps S11 and S12 to obtain the initial embedding vector. It is important to note that q is not an existing micro-creative resource on the blockchain; it lacks a unique identifier and historical propagation path. Therefore, no propagation neighbors are constructed; the initial embedding vector is generated solely using the content modality.

[0163] The initial query embedding vector is input into the shared projection layer trained in step S41 and the modality-specific projection module corresponding to step S42 to obtain the representation in the query unified semantic space. , dimension (Example: 128). Read the representation of each candidate micro-creative resource c in the unified semantic space from the candidate micro-creative resource set. The candidate micro-creative resource set can include all micro-creative resources that can currently be displayed or recommended on the platform.

[0164] calculate and Calculate the cosine similarity between them, and select the top two with the highest similarity. A preliminary recall set was formed from the candidate micro-creative resources. This indicates the recall quantity, with examples of 100 or 500. If query q has no historical propagation record, it can still enter the unified semantic space through content encoding and modality-specific projection, thereby supporting the recommendation of new creative needs and micro-creative resources on the blockchain.

[0165] S62, calculate the recommendation score based on the cosine similarity and the predicted influence of the candidate resource. The recommendation score is the sum of the cosine similarity and the first product, where the first product is the product of the cosine similarity, the preset coefficient, and the predicted influence of the candidate resource.

[0166] For each candidate micro-creative resource c in the initial recall set, the head of propagation influence is predicted. Calculate predictive impact , The value ranges from 0 to 1 and is used to characterize the potential dissemination influence of candidate micro-creative resource c. If candidate micro-creative resource c has historical dissemination popularity weights... If the popularity score is not recorded, both the popularity weight and the predicted influence will be retained; if there is no popularity record, only the predicted influence will be output as a cold start reference.

[0167] Calculate the recommendation score between the creative demand query q and the candidate micro-creative resources c. , This represents the combined recommendation score of the creative demand query q and the candidate micro-creative resource c. The higher the score, the more suitable the candidate micro-creative resource c is to be recommended to the creative demand query q.

[0168] In one implementation, the representation in the query unified semantic space is first computed. Representation in the unified semantic space of candidate micro-creative resources Cosine similarity between This is used to characterize the semantic similarity between the two in a unified semantic space. Then, a lightweight weighting is applied using the predicted propagation influence of the candidate micro-creative resources to obtain the recommendation score:

[0169] in, This is the propagation influence bias coefficient, used to control the strength of the predicted propagation influence on the ranking results; its value ranges from 0.1 to 0.3. If... If the size is too large, the ranking results may be overly biased towards micro-creative resources with high predictive dissemination influence; if If the value is too small, the predicted dissemination influence will not contribute sufficiently to the ranking results.

[0170] This represents the predicted dissemination impact of candidate micro-creative resource c, i.e., the dissemination impact prediction head. Based on the representation in the unified semantic space of micro-creative resource c The output normalized predicted value ranges from [0,1]. It is obtained by regression training on micro-creative resources with historical dissemination popularity records, and is used to estimate the potential dissemination influence of a micro-creative resource when there are no real popularity records (cold-start micro-creative resources) or when comprehensive reference is required.

[0171] S63. Sort the candidate resources from highest to lowest recommendation score to obtain the recommendation results.

[0172] According to the recommended score The candidate micro-creative resources in the initial recall set are sorted from highest to lowest, and a final recommendation list is output. Each candidate micro-creative resource in the final recommendation list includes at least: a unique identifier for the micro-creative resource, its primary modality, a text summary, an image thumbnail or audio preview, cosine similarity, predicted influence, and a recommendation score.

[0173] Optionally, the final recommendation list can be filtered: if the user account specifies a target modality (e.g., only audio is needed), candidate micro-creative resources that do not contain audio are filtered out; if the user account specifies a lower limit for popularity, candidate micro-creative resources whose predicted influence is below the threshold are filtered out; if the user account specifies an authorization type, candidate micro-creative resources that do not meet the authorization type requirements are filtered out.

[0174] It also outputs recommendation explanations, including reasons for content matching and reasons for influence. Reasons for content matching can come from text keywords, image style tags, or audio style tags; reasons for influence can come from predicted influence, historical dissemination popularity, recent attention frequency, and graph neighbor strength.

[0175] In this embodiment, the user account inputs a text query "deep ambient sound for sci-fi short film opening". The system encodes it into a representation in the unified semantic space for queries . The candidate micro-creative resource library contains micro-creative resource A (city night scene audio of a specific style, with a predicted influence of 0.63), micro-creative resource E (an unpropagated cold-start micro-creative resource, with a predicted influence of 0.45), and micro-creative resource F (sci-fi style background music, with a predicted influence of 0.5). The cosine similarity is calculated as: 0.82 with micro-creative resource A, 0.78 with micro-creative resource E, and 0.75 with micro-creative resource F. Take =0.2, then the recommendation scores are: micro-creative resource A = 0.82×(1+0.2×0.63) = 0.82×1.126 ≈ 0.923; micro-creative resource E = 0.78×(1+0.09) = 0.78×1.09 ≈ 0.850; micro-creative resource F = 0.75×(1+0.1) = 0.75×1.1 = 0.825. The final ranking is micro-creative resource A, micro-creative resource E, and micro-creative resource F. Micro-creative resource A obtains a higher ranking due to its high predicted propagation influence, and micro-creative resource E can still be recommended despite having no propagation influence record due to its high content matching degree. When outputting the recommendation list, an explanation is attached for micro-creative resource A: "Content matching: The deep ambient sound style is similar to the specific style sound effect; Propagation influence reference: This micro-creative resource has a recent access heat of 2500 times, with a high predicted propagation influence."

[0176] In one embodiment, as shown in Figure 7 , the horizontal bar chart shows the recommendation scores of the top 15 candidate micro-creative resources returned for a certain creative demand query. The horizontal axis is "recommendation score" (dimensionless), and the vertical axis is the unique identifier of the micro-creative resource. Each bar is marked with the semantic similarity value and the predicted propagation influence value. The recommendation score is obtained by multiplying the cosine similarity by the propagation influence bias, and the ranking is from high to low, which proves that the recommendation mechanism can make candidates with high propagation influence obtain higher scores on the premise of ensuring the priority of semantic matching.

[0177] As an optional implementation step, after step S6, step S7 is further included: when a new micro-creative resource or a new historical propagation path is added, the propagation directed graph is updated, and graph neural network propagation is re-performed on the affected micro-creative resources, so as to update the graph-enhanced micro-creative resource vector, the representation in the unified semantic space, and the propagation influence prediction branch.

[0178] S71: when a new micro-creative resource appears, generate its initial embedding vector according to the steps of collection and standardization.

[0179] When a new micro-creative resource is added, step S11 is followed to collect the text, image, audio, and historical propagation path of the new micro-creative resource stored on the blockchain, and step S12 is followed to generate its initial embedding vector. If the new micro-creative resource has not yet been recorded for attention, the micro-creative resource will only enter the directed graph of micro-creative resource propagation through self-loop edges.

[0180] S72, when a new historical propagation path appears, update the propagation trajectory of the address of interest and update the edge parameters of the directed graph.

[0181] When a new historical dissemination path appears, the user account's follow history is updated based on the user account, timestamp, and unique identifier of the micro-creative resource. If the same user account forms a new continuous follow relationship within a preset time window, a new directed edge is established or the repeated follow count of an existing directed edge is updated according to step S2. Attention time interval and edge weight .

[0182] S73, re-perform graph attention propagation on the affected micro-creative resource vertices and their neighboring vertices, and update the graph-enhanced micro-creative resource vectors and their representations in the unified semantic space.

[0183] For micro-creative resource vertices and their neighboring vertices affected by the newly added historical propagation paths, the graph attention propagation based on heat-time behavior bias in steps S31 to S34 is re-executed to update the corresponding graph-enhanced micro-creative resource vectors and their representations in the unified semantic space. If computational resources are sufficient, the directed graph of the propagation of all micro-creative resources can be recalculated periodically; if computational resources are limited, only the vertices at the two ends of the newly added directed edges and all vertices reachable from these vertices in at most two steps (i.e., two edges) can be locally updated. Among them, the update of the two ends of the vertices themselves will affect their direct neighbors (1 hop), and the change of the 1-hop neighbors will further affect the vertices directly connected to them (2 hops), but the influence beyond 2 hops has been significantly weakened and can be temporarily left unupdated to save computational resources.

[0184] S74 calculates the predicted influence of newly added micro-creative resources through the dissemination influence prediction branch and makes incremental fine-tuning according to a preset cycle.

[0185] Calculate and predict the impact of newly added micro-creative resources The newly added micro-creative resources will be added to the candidate micro-creative resource set. If the newly added micro-creative resources subsequently generate real dissemination popularity records, they will be added to the micro-creative resource set with historical dissemination popularity records in the next round of training or incremental fine-tuning. It is used to update the propagation influence prediction header.

[0186] The model is incrementally fine-tuned according to a preset period, which can be daily, weekly, or every 1000 newly added historical dissemination paths. During incremental fine-tuning, priority is given to sampling recently added cross-modal positive sample micro-creative resource pairs and recently added dissemination influence perception level triples, so that the unified semantic space can adapt to new dissemination trends.

[0187] In this embodiment, if the micro-creative resource E has multiple access records after a period of time (e.g., the cumulative number of visits reaches 1500), the system will perform an incremental update: updating the micro-creative resource E's... The normalized propagation influence and graph augmentation vector are recalculated and added to the propagation influence prediction head training in the next round of fine-tuning, making the predicted influence of micro-creative resource E closer to the true level. At the same time, if a new user account continuously follows micro-creative resource E and another new micro-creative resource G, a new directed edge is established and the graph is updated. The affected vertices re-perform graph attention propagation, so that the micro-creative resource representation can reflect recent propagation changes in a timely manner.

[0188] Example 2

[0189] This embodiment provides a blockchain-based micro-creative resource semantic understanding and intelligent recommendation system for implementing the method described in Embodiment 1. The system includes:

[0190] 1. Data acquisition module, used to acquire multimodal content data of micro-creative resources and their historical propagation path on the blockchain, and generate initial embedding vectors.

[0191] Specifically, this module standardizes the text, images, and audio of micro-creative resources, groups and sorts historical dissemination paths by user account, and outputs initial embedding vectors.

[0192] 2. A graph propagation enhancement module is used to construct a propagation directed graph with micro-creative resources as vertices based on the historical propagation path, and to integrate the propagation behavior features into the vertex representation through a graph neural network to obtain a graph-enhanced micro-creative resource vector.

[0193] Specifically, this module establishes directed edges between vertices based on the continuous follow trajectory of user accounts, calculates edge weights, and performs graph attention propagation to integrate information such as propagation popularity, time freshness, and number of repeated follows into vertex features. Finally, it outputs a graph-enhanced micro-creative resource vector for use by the subsequent cross-modal alignment module.

[0194] 3. A cross-modal alignment module is used to map the graph-enhanced micro-creative resource vectors to a unified semantic space through contrastive learning.

[0195] Specifically, the module searches for micro-creative resource pairs with different main modalities as positive samples within a preset time window. It maps the graph augmentation vectors to representations in a unified semantic space through a shared projection layer and a modality-specific compensation layer, and trains them using temperature-scaled contrastive loss. Finally, it outputs the representation of each micro-creative resource in the unified semantic space for subsequent influence constraint training and recommendation ranking.

[0196] 4. Influence Constraint Training Module: This module is used to introduce propagation influence constraints into the unified semantic space for model training, and to add a propagation influence prediction branch to obtain the predicted influence of the micro-creative resource.

[0197] Specifically, this module divides the propagation influence into high, medium and low influence sets based on normalized propagation influence, constructs propagation influence perception triples, and uses dynamic margin contrast loss and high influence cluster compact regularization for training. At the same time, it outputs predicted influence through prediction head.

[0198] 5. The recommendation and incremental update module is used to receive queries, encode the queries and map them to the unified semantic space, perform similarity matching with candidate micro-creative resources, combine the predicted influence to obtain recommendation results, and perform incremental updates when new micro-creative resources or new historical propagation paths are added.

[0199] Specifically, this module receives text, image, or audio queries input by the user, encodes them, maps them to a unified semantic space, calculates the similarity with candidate micro-creative resources, and combines the predicted influence to obtain a recommendation score and output it in a ranking. When a new micro-creative resource or a new historical propagation path is added, this module updates the propagation directed graph, re-executes graph attention propagation on the affected vertices, and updates the graph to enhance the micro-creative resource vector, the representation in the unified semantic space, and the propagation influence prediction branch.

[0200] The above modules are connected sequentially to jointly realize a complete set of functions, from data acquisition, graph enhancement, cross-modal alignment, influence clustering training to intelligent recommendation and incremental updates. For details on the specific implementation of the operations performed by each module, please refer to the relevant steps described in Example 1; they will not be repeated here.

[0201] Example 3

[0202] This embodiment provides an electronic device, the electronic device comprising: A memory for storing computer programs, the memory including non-volatile memory and random access memory, the non-volatile memory for storing system programs and the random access memory for caching runtime data; The processor is used to execute the computer program stored in the memory to implement the steps of the blockchain-based micro-creative resource semantic understanding and intelligent recommendation method as described in Embodiment 1.

[0203] For details on the specific implementation of each step and related explanations, please refer to Example 1, which will not be repeated here.

[0204] The processor can be a general-purpose processor, including a central processing unit, a network processor, etc.; it can also be a digital signal processor, an application-specific integrated circuit, a field-programmable gate array, or other programmable logic devices.

[0205] Example 4

[0206] This embodiment provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the steps of a blockchain-based micro-creative resource semantic understanding and intelligent recommendation method as described in Embodiment 1.

[0207] For details on the specific implementation of each step of this method and related explanations, please refer to Example 1, which will not be elaborated here.

[0208] The computer-readable storage medium can be any tangible medium that contains or stores program instructions, such as: USB flash drive, portable hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk or optical disk, etc.

[0209] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A blockchain-based method for semantic understanding and intelligent recommendation of micro-creative resources, characterized in that, include: Acquire multimodal content data of micro-creative resources to generate an initial embedding vector, wherein the multimodal content data includes one or more of text, images, and audio; For any micro-creative resource in the micro-creative resource set, its blockchain account identifier, work number, text, image, audio, and historical propagation path stored on the blockchain are read. Based on the historical propagation path of the micro-creative resource on the blockchain, a propagation directed graph is constructed with the micro-creative resource as the vertex. The initial embedding vector is used as the initial feature. The propagation behavior features are integrated into the vertex representation through a graph neural network to obtain a graph-enhanced micro-creative resource vector. By contrastive learning, the graph-enhanced micro-creative resource vectors are mapped to a unified semantic space, and a propagation influence constraint is introduced into the unified semantic space. Then, a propagation influence prediction branch is added to output the predicted propagation influence corresponding to the micro-creative resource based on the representation of the micro-creative resource in the unified semantic space. The training process for the unified semantic space and the propagation influence prediction branch includes: The primary modality of each micro-creative resource is determined, and pairs of micro-creative resources that are followed sequentially within a preset time window and have different primary modalities are identified as cross-modal positive sample pairs. The cross-modal positive sample pairs are used as positive samples for contrastive learning, and the two representations corresponding to the cross-modal positive sample pairs are brought closer in the unified semantic space to obtain the cross-modal alignment loss. Based on normalized dissemination influence, micro-creative resources are divided into sets of high dissemination influence micro-creative resources, sets of medium dissemination influence micro-creative resources, and sets of low dissemination influence micro-creative resources. Anchor points are selected from the high or medium influence sets, positive samples are selected from resources with the same or higher influence level as the anchor points, and negative samples are selected from resources with a lower influence level than the anchor points to form triples. The contrast loss margin of the triplet is adjusted based on the difference in the influence of the triplets. The contrast loss of the triplet is expressed as follows: Where μ represents the basic margin hyperparameter, This represents the embedding distance between anchor micro-creative resource a and positive sample micro-creative resource b. This represents the embedding distance between the anchor micro-creative resource a and the negative sample micro-creative resource e; The dynamic adjustment item is , , , These represent anchor point micro-creative resource a and positive sample micro-creative resource, respectively. Negative Sample Micro-Creative Resources The influence of normalized dissemination; This indicates the operation of retrieving the maximum value; This represents the moderating strength coefficient for differences in dissemination influence; ε represents a minimum positive number to prevent the denominator from being zero. Apply compact regularization constraints to the representations of micro-creative resources in the high-impact set, calculate the cluster center of high-propagation-impact micro-creative resources, then calculate the distance between the representation of micro-creative resources in the unified semantic space of each high-propagation-impact micro-creative resource and the cluster center of high-propagation-impact, and average these distances as the high-propagation-impact compact regularization term; Combining all triple contrast loss and high propagation influence compact regularization term yields propagation influence clustering loss; The dissemination influence prediction branch is trained using micro-creative resources with historical dissemination popularity. Based on the predicted dissemination influence and normalized dissemination influence of the micro-creative resources, the dissemination influence regression loss is calculated. The normalized dissemination influence is calculated based on historical dissemination popularity. The cross-modal alignment loss, the propagation influence clustering loss, and the propagation influence regression loss are combined into a joint optimized total loss, and then trained in multiple rounds. The system receives a query, encodes the query, maps it to a trained unified semantic space to obtain a query representation, selects candidate resources from the micro-creative resource set, obtains the candidate resource representation mapped to the trained unified semantic space, performs similarity matching between the query representation and the candidate resource representation, and combines the predicted propagation influence of the candidate resources output by the trained propagation influence prediction branch to obtain a recommendation result.

2. The method according to claim 1, characterized in that, The generation of the initial embedding vector includes: Each modality in the multimodal content data is encoded separately to obtain the initial features of each modality; Construct a modal effective vector, wherein each component of the modal effective vector corresponds one-to-one with each mode; The initial features of each modality are input into the modality projection layer corresponding to that modality to obtain the modality projection features; Based on the modality effective vector, the gating score corresponding to the missing modality is assigned a minimum value, and the gating score of the existing modality is normalized to obtain the gating weight. The initial embedding vector is obtained by weighting and summing the projection features of each modality according to the gating weights.

3. The method according to claim 1, characterized in that, The construction of the propagation directed graph and the integration of propagation behavior features into the vertex representation through a graph neural network to obtain the graph-enhanced micro-creative resource vector include: Using the micro-creative resources as vertices, the historical propagation paths are sorted chronologically by user accounts. Directed edges are established between adjacent vertices and the number of repeated attentions is accumulated. Edge weights are calculated based on the number of repeated attentions, the propagation popularity weights in the historical propagation paths, and the attention time intervals in the historical propagation paths. Self-loop edges are added and adjacent vertices whose weights meet preset conditions are retained. Construct a feature vector containing the edge weights and the directions of the directed edges, map it to a behavior bias and superimpose it on the content relevance score, normalize it to obtain an attention coefficient, and aggregate the neighbor vertex features according to the weight of the coefficient to update the center vertex features, thus obtaining a graph-enhanced micro-creative resource vector.

4. The method according to claim 1, characterized in that, The method of mapping graph-enhanced micro-creative resource vectors to a unified semantic space through contrastive learning includes: The graph-enhanced micro-creative resource vectors are input into the shared projection layer to obtain the shared basic vectors; Select the corresponding modality-specific projection module based on the main modality of the micro-creative resource, and output the modality compensation vector; The shared fundamental vector and the modal compensation vector are added together, and the result is layer-normalized to obtain a representation of micro-creative resources in a unified semantic space.

5. The method according to claim 1, characterized in that, The predicted propagation influence of the candidate resources, combined with the output of the propagation influence prediction branch after training, yields the recommendation result, including: Calculate the cosine similarity between the query representation and the candidate resource representation; A recommendation score is calculated based on the cosine similarity and the predicted propagation influence of the candidate resource. The recommendation score is the sum of the cosine similarity and a first product, where the first product is the product of the cosine similarity, a preset coefficient, and the predicted propagation influence of the candidate resource. Candidate resources are sorted from highest to lowest according to their recommendation scores to obtain the recommendation results.

6. The method according to claim 1, characterized in that, Also includes: When a new micro-creative resource or a new historical propagation path is added, the propagation directed graph is updated and the graph neural network propagation is re-executed on the affected micro-creative resources, updating the graph-enhanced micro-creative resource vector, the representation of the micro-creative resource in the unified semantic space, and the propagation influence prediction branch.

7. A blockchain-based micro-creative resource semantic understanding and intelligent recommendation system, used to implement the method as described in any one of claims 1-6, characterized in that, include: The data acquisition module is used to acquire multimodal content data of micro-creative resources and their historical propagation paths on the blockchain, and generate an initial embedding vector; The graph propagation enhancement module is used to construct a propagation directed graph with micro-creative resources as vertices based on the historical propagation path, and to integrate the propagation behavior features into the vertex representation through a graph neural network to obtain a graph-enhanced micro-creative resource vector. A cross-modal alignment module is used to map the graph-enhanced micro-creative resource vectors to a unified semantic space through contrastive learning; The influence constraint training module is used to introduce propagation influence constraints into the unified semantic space for model training, and to add a propagation influence prediction branch to obtain the predicted propagation influence of the micro-creative resources. The recommendation and incremental update module is used to receive queries, encode the queries and map them to the unified semantic space, perform similarity matching with candidate micro-creative resources, combine the predicted dissemination influence to obtain recommendation results, and perform incremental updates when new micro-creative resources or new historical dissemination paths are added.

8. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing a computer program stored in the memory to implement the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the method according to any one of claims 1 to 6.

Images