A data processing method and system for a digital copyright evaluation model

Abstract

This invention relates to the field of digital copyright assessment technology, and discloses a data processing method and system for a digital copyright assessment model. The method includes: acquiring multi-source heterogeneous data; constructing a dynamic knowledge graph; performing time-series processing on the dynamic knowledge graph to generate several time-series graph snapshots; extracting the number of dissemination platforms, the rate of change of core features, and the assessment target labels based on the time-series graph snapshots, and selecting a spatiotemporal graph neural network model; inputting the time-series graph snapshots into the selected spatiotemporal graph neural network model for spatiotemporal feature learning to generate spatiotemporal feature vectors; performing feature fusion processing on the spatiotemporal feature vectors and inputting them into the copyright value assessment model to obtain the digital copyright value assessment result. This application improves the adaptability and assessment accuracy to complex dynamic scenarios.

Description

Technical Field

[0001] This invention relates to the field of digital copyright assessment technology, and more specifically, to a data processing method and system for digital copyright assessment models. Background Technology

[0002] With the development of the digital content industry, the dissemination of digital copyrights has gradually shifted from single-media dissemination to multi-platform, multi-entity interactive dissemination. The value of digital copyrights is comprehensively influenced by various factors, including the influence of the dissemination platform, changes in user behavior, the scope of the work's dissemination, changes in market popularity, and adjustments in the policy environment. Therefore, accurately acquiring the dynamic characteristics affecting the value of digital copyrights and achieving reliable assessment based on these dynamic changes has become a crucial technical issue in the management and transaction of digital copyrights.

[0003] Current digital copyright valuation methods typically employ a single, fixed feature extraction model to uniformly process all copyrighted works. This "one-size-fits-all" fixed model architecture cannot adapt to the dynamic differences in factors influencing digital copyright value across different scenarios (such as the drastic change from "stable dissemination on a single platform" to "explosive dissemination across multiple platforms"), leading to imbalances in the following two dimensions: Imbalance between accuracy and resources: While high-complexity models (such as the Transformer architecture) can capture complex relationships, they incur enormous computational costs; using high-complexity models uniformly for all works (including low-fluctuation existing works) results in a waste of computational resources; conversely, using lightweight models for highly dynamic hit works obscures key features (such as hot topics and cross-platform collaborations), severely reducing valuation accuracy. Lack of model evolution capabilities: Even when existing methods employ machine learning models, parameter updates rely on periodic full retraining by humans, failing to automatically trigger incremental fine-tuning of the model based on real-time dynamic data (such as sudden changes in dissemination popularity or the release of new policies and regulations), resulting in a disconnect between model representation and the current market state.

[0004] Therefore, it is necessary to design a data processing method and system for digital rights assessment models to solve the problems existing in the current technology. Summary of the Invention

[0005] In view of this, the present invention proposes a data processing method and system for digital copyright assessment models, aiming to solve the problem that existing digital copyright value assessments cannot adapt to the dynamic differences of digital copyright value influencing factors in different scenarios, which easily leads to an imbalance between accuracy and resources and a lack of model evolution capabilities.

[0006] In one aspect, the present invention proposes a data processing method for a digital rights assessment model, comprising:

[0007] Acquire multi-source heterogeneous data; the multi-source heterogeneous data includes copyrighted work data, dissemination platform data, user behavior data, and legal entity data;

[0008] Copyrighted work data, dissemination platform data, user behavior data, and legal entity data are transformed into entity nodes, and relationship edges are constructed based on the dissemination relationship, consumption relationship, authorization relationship, and association relationship between the entity nodes to form a dynamic knowledge graph;

[0009] The dynamic knowledge graph is processed in a time sequence based on a preset time window to generate several time-series graph snapshots.

[0010] The number of propagation platforms, the rate of change of core features, and the target labels are extracted based on the time series graph snapshot, and a spatiotemporal graph neural network model is selected. The time series graph snapshot is then input into the selected spatiotemporal graph neural network model to learn spatiotemporal features and generate spatiotemporal feature vectors.

[0011] The spatiotemporal feature vectors are subjected to feature fusion processing and input into the copyright value assessment model to obtain the digital copyright value assessment results.

[0012] Furthermore, the entity nodes include copyrighted work nodes, dissemination platform nodes, user nodes, and legal entity nodes, and each entity node is configured with basic attributes and dynamic attributes;

[0013] The relationship edges include dissemination relationship edges, consumption relationship edges, licensing relationship edges, and association relationship edges; the dissemination relationship edge points from the copyrighted work node to the dissemination platform node; the consumption relationship edge points from the user node to the copyrighted work node; the licensing relationship edge points from the copyrighted work node to the legal entity node; and the association relationship edge points from the dissemination platform node to the user node.

[0014] Furthermore, the dynamic knowledge graph adopts a three-layer storage architecture of nodes-edges-attributes, including node storage, edge storage, and attribute storage.

[0015] Furthermore, the dynamic knowledge graph is dynamically updated, including dynamic updates of node attributes, dynamic updates of edge relationships, and incremental generation and updates of graph snapshots.

[0016] Furthermore, when selecting a spatiotemporal graph neural network model, the following factors are considered:

[0017] Extract the number of dissemination platforms for copyrighted works, the rate of change of core features within a preset period, and the evaluation target label from the time series graph snapshot; compare the number of dissemination platforms, the rate of change of core features, and the evaluation target label with the basic scene rules, core scene rules, and complex scene rules respectively to determine the scene type; based on the determined scene type, select the matching spatiotemporal graph neural network model from ST-GCN (Spatial-Temporal Graph Convolutional Network), AST-GNN (Attention-based Spatial-Temporal Graph Neural Network), and ST-TransGNN (Spatial-Temporal Transformer Graph Neural Network);

[0018] The system monitors the rate of change of the core features and the number of propagation platforms in real time, and triggers the adaptive switching of the spatiotemporal graph neural network model when the scenario upgrade conditions are met.

[0019] Furthermore, the generation of the spatiotemporal feature vector includes: when the selected spatiotemporal graph neural network model is the ST-GCN model, a short-time snapshot sequence is selected as input data; in the spatial feature learning stage, a graph convolution message passing mechanism is adopted, with the copyrighted work node as the central node, the features of 1-hop neighbor nodes are aggregated, and the spatial feature vector is output; in the temporal feature learning stage, the temporal pattern within a short period of the copyrighted work is captured by one-dimensional convolution, and the temporal feature vector is output; the spatial feature vector and the temporal feature vector are concatenated at a fixed ratio to generate the first-dimensional spatiotemporal feature vector.

[0020] Furthermore, the generation of the spatiotemporal feature vector also includes: when the selected spatiotemporal graph neural network model is an AST-GNN model, selecting an incremental temporal snapshot sequence as input data; in the spatial feature learning stage, using a graph attention network message passing mechanism, calculating the attention coefficients of the copyrighted work node and its neighboring nodes and assigning weights, outputting the spatial feature vector; in the temporal feature learning stage, using a causal convolution combined with a temporal attention mechanism, assigning weights based on the rate of change of the core features, outputting the temporal feature vector; and using multi-head attention weighted fusion and residual connections to fuse the spatial feature vector and the temporal feature vector to generate the second-dimensional spatiotemporal feature vector.

[0021] Furthermore, the generation of spatiotemporal feature vectors also includes: when the selected spatiotemporal graph neural network model is the ST-TransGNN model, a long-term snapshot sequence is selected as input data, and a graph sampling aggregation algorithm is used to sample and aggregate node features; in the spatial feature learning stage, a self-attention mechanism is used to calculate the long-distance correlation weights between nodes, capture cross-platform propagation linkage features, and output the spatial feature vector; in the temporal feature learning stage, a Transformer temporal encoder is used to capture the feature evolution law within a long period through a self-attention mechanism, and output the temporal feature vector; a cross-spatiotemporal self-attention fusion layer is used to fuse the spatial feature vector and the temporal feature vector to generate the third-dimensional spatiotemporal feature vector.

[0022] Furthermore, it also includes incremental training of the spatiotemporal graph neural network model:

[0023] Incremental training is triggered when the number of newly added snapshots in the time series map exceeds a certain threshold.

[0024] Incremental training is triggered when the absolute value of the rate of change of the core features of key nodes within a preset period exceeds the rate of change threshold.

[0025] Based on the newly added time series graph snapshot, the parameters of the spatiotemporal graph neural network model near the output layer are adjusted, and the parameters of the spatiotemporal graph neural network model are updated.

[0026] Compared with existing technologies, the advantages of this invention are as follows: The dynamic knowledge graph integrates multi-source heterogeneous entities such as copyrighted works, dissemination platforms, users, and legal entities, breaking through the limitations of traditional evaluation that only focuses on the work itself and capturing the interactive influence of related entities; ST-GNN realizes the joint learning of spatial structural relationships and time series features, and the generated spatiotemporal feature representation vectors can better reflect the dynamic changes in copyright value, with stronger representation capabilities compared to single static features or time series features; The automatic update engine realizes full automation of data collection, graph update, and model evolution without manual intervention, ensuring that feature data and model representations are always up-to-date and adaptable to dynamic changes such as market trends and policy adjustments; Comprehensive, dynamic, and deep feature representations provide higher quality input for subsequent evaluation models, improving the evaluation model's adaptability to complex dynamic scenarios and its evaluation accuracy.

[0027] On the other hand, this application also provides a data processing system for a digital rights management (DRM) assessment model, used to apply the above-mentioned data processing method for a DRM assessment model, including:

[0028] The data acquisition module is configured to acquire multi-source heterogeneous data, including copyrighted work data, dissemination platform data, user behavior data, and legal entity data.

[0029] The dynamic knowledge graph construction module is configured to transform copyrighted work data, dissemination platform data, user behavior data, and legal entity data into entity nodes, and construct relationship edges based on the dissemination relationship, consumption relationship, authorization relationship, and association relationship between the entity nodes to form a dynamic knowledge graph.

[0030] The snapshot generation module is configured to perform time-series processing on the dynamic knowledge graph based on a preset time window to generate several time-series graph snapshots.

[0031] The model adaptive selection module is configured to extract the number of propagation platforms, the rate of change of core features, and the evaluation target label based on the time series graph snapshot, and select a spatiotemporal graph neural network model; and input the time series graph snapshot into the selected spatiotemporal graph neural network model to perform spatiotemporal feature learning and generate spatiotemporal feature vectors.

[0032] The digital copyright assessment module is configured to perform feature fusion processing on the spatiotemporal feature vector and input it into the copyright value assessment model to obtain the digital copyright value assessment result.

[0033] It is understandable that the data processing methods and systems used in the digital rights assessment model described above have the same beneficial effects, and will not be elaborated further here. Attached Figure Description

[0034] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0035] Figure 1 A flowchart of a data processing method for a digital rights assessment model provided in an embodiment of the present invention;

[0036] Figure 2 This is a functional block diagram of a data processing system for a digital rights assessment model provided in an embodiment of the present invention. Detailed Implementation

[0037] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0038] In some embodiments of this application, see Figure 1 As shown, a data processing method for a digital rights assessment model is proposed, including:

[0039] S100: Acquire multi-source heterogeneous data; multi-source heterogeneous data includes copyrighted work data, distribution platform data, user behavior data, and legal entity data;

[0040] S200: Transform copyrighted work data, dissemination platform data, user behavior data, and legal entity data into entity nodes, and construct relationship edges based on the dissemination relationship, consumption relationship, authorization relationship, and association relationship between each entity node to form a dynamic knowledge graph;

[0041] S300: Performs time-series processing on dynamic knowledge graphs based on preset time windows to generate several time-series graph snapshots;

[0042] S400: Extract the number of propagation platforms, the rate of change of core features, and the evaluation target labels based on the time series graph snapshot, and select a spatiotemporal graph neural network model; input the time series graph snapshot into the selected spatiotemporal graph neural network model to learn spatiotemporal features and generate spatiotemporal feature vectors;

[0043] S500: Perform feature fusion processing on spatiotemporal feature vectors and input them into the copyright value assessment model to obtain the digital copyright value assessment results.

[0044] In step S100, to achieve a comprehensive understanding of the factors influencing the value of digital copyright, multi-source heterogeneous data is first acquired during the data processing for the digital copyright assessment model. This multi-source heterogeneous data includes copyrighted work data, dissemination platform data, user behavior data, and legal entity data. The multi-source heterogeneous data is used to characterize information across multiple dimensions, such as the digital copyright object's own attributes, dissemination environment, user feedback, and the status of copyright-related entities.

[0045] Specifically, copyright work data is used to describe the basic information and historical changes of the digital copyright object itself, including but not limited to work identification information, work name, work type, creation time, publication time, content classification information, copyright registration information, historical dissemination records, and historical value assessment records. Among them, work identification information is used to uniquely identify the digital copyright object to be evaluated; work type information is used to distinguish the dissemination characteristics corresponding to different categories of digital content; historical dissemination records are used to reflect the dissemination trend of the copyright work in different time periods; and historical value assessment records are used to provide reference data for subsequent model training and feature analysis.

[0046] Distribution platform data describes the dissemination status of digitally copyrighted works across different distribution channels, including but not limited to platform type, platform identification information, work display location, number of plays, number of views, number of likes, number of comments, number of shares, number of favorites, dissemination chain information, and platform user scale information. The distribution platform data interacts with various digital content distribution platforms through configured data collection interfaces, including but not limited to platform open interfaces, data exchange interfaces, and data synchronization interfaces. By utilizing distribution platform data, it is possible to obtain the scope, speed, and platform influence of copyrighted works across different distribution platforms, providing data support for subsequent analysis of the impact of changes in copyright value with the dissemination environment.

[0047] User behavior data is used to describe the interaction between users and digital copyrighted works, including but not limited to user access records, viewing time, search behavior, click behavior, collection behavior, comment behavior, sharing behavior, and information on changes in user groups. Specifically, during the collection process, user interaction data, after being anonymized, is obtained through a user behavior tracking system. Corresponding data relationships are established based on user identifiers, work identifiers, and the time of the behavior to reflect the level of attention, consumption preferences, and dissemination impact of different user groups on digital copyrighted works.

[0048] Legal entity data is used to describe information related to the rights holders and legal environment associated with digital copyright, including but not limited to copyright holder information, creator information, operating entity information, licensing relationship information, copyright transaction information, legal entity qualification information, and policy and regulatory information. Specifically, legal entity data is obtained by connecting to policy release platforms, copyright registration databases, and licensing management systems, and is used to reflect the legal relationships, ownership status, and changes in the external policy environment corresponding to digital copyright.

[0049] In acquiring the aforementioned multi-source heterogeneous data, to ensure that the data acquisition process can adapt to the dynamic changes in factors affecting digital copyright value, this embodiment classifies the acquired data features based on their degree of influence on the copyright value assessment results and their own frequency of change, forming core feature data and non-core feature data.

[0050] Core feature data represents dynamic data that changes frequently and significantly impacts copyright value assessment results. This includes real-time dissemination data, user interaction data, and public opinion data. For example, real-time play counts, number of reposts, comment popularity, and user activity directly reflect the current dissemination status of digital copyrights. Their trends typically have a rapid impact on copyright value, hence the high data collection frequency. The importance of core feature data is determined based on its own volatility and its sensitivity to the output of the copyright value assessment model. Specifically, the stability of data feature changes is calculated by statistically analyzing the magnitude of data changes within a preset time period. Simultaneously, the influence weight of different data features on the assessment results is determined by evaluating the contribution of features in the model. When a data feature exhibits significant volatility and its impact on the assessment results exceeds preset conditions, it is identified as core feature data.

[0051] Non-core feature data is used to represent data with long change cycles that will not significantly affect the copyright value assessment results in the short term. This includes platform user scale, legal entity qualifications, and basic information of the copyright holder. For example, the number of registered users on a distribution platform, basic information of the copyright holder, and the authentication status of the legal entity usually do not change significantly in a short period of time, so there is no need to use high-frequency collection methods. For this type of data, its latest status is obtained through periodic collection to reduce the consumption of data collection resources.

[0052] Data collection frequency is dynamically configured based on the importance of different features. For core feature data, higher-frequency data collection tasks are configured, such as collecting real-time dissemination data and public opinion data on an hourly basis, for example, once per hour, to ensure timely capture of sudden changes in the digital copyright dissemination process. For non-core feature data, lower-frequency data collection tasks are configured, such as collecting platform user scale and legal entity qualification information on a daily basis, for example, once every 24 hours, to reduce the data processing burden caused by repeated collection. In some implementations, the collection cycle can be dynamically adjusted according to historical data change trends. When an abnormal change trend is detected in a non-core feature, the collection frequency of the corresponding data source is increased, temporarily transforming it into key monitoring data.

[0053] After collecting data from different sources, the acquired data undergoes unified preprocessing, including data format conversion, field mapping, timestamp synchronization, anomaly filtering, and data integrity verification. Specifically, to address the differences in data formats across different data sources, the data is converted into a unified data structure. To address time discrepancies between different systems, time synchronization is performed based on the data's generation time, ensuring that copyrighted works, distribution platforms, user behavior, and legal entity data can be correlated under the same time benchmark. Missing, duplicate, and anomaly data are identified and processed using pre-defined data verification rules to guarantee the accuracy of the data input into the subsequent dynamic knowledge graph construction process.

[0054] In step S200, after acquiring and preprocessing multi-source heterogeneous data, copyrighted work data, dissemination platform data, user behavior data, and legal entity data are converted into entity nodes in a dynamic knowledge graph. Corresponding relationship edges are then constructed based on the business interaction relationships, data association relationships, and temporal evolution relationships between different entity nodes to form a dynamic knowledge graph describing the digital copyright dissemination ecosystem. The dynamic knowledge graph is used to connect and integrate copyright object information, dissemination subject information, user interaction information, and legal service information that were originally scattered across different data sources, thereby forming a data organization structure that reflects the factors influencing the value of digital copyright and its dynamic changes.

[0055] Specifically, the dynamic knowledge graph includes four core entity nodes: copyrighted works nodes, dissemination platform nodes, user nodes, and legal entity nodes.

[0056] The copyright work node represents a digital copyright object to be evaluated. Each copyright work node corresponds to a unique digital copyright object and is configured with basic attributes describing the work's inherent properties and dynamic attributes describing its dynamic changes during dissemination. Basic attributes include work identification information, work type information, work content feature information, and publication time information. Specifically, work identification information is used to uniquely identify different copyrighted works; work type information is used to distinguish different types of digital copyright objects such as text works, image works, audio-visual works, and short video works; work content feature information represents the work's own content characteristics, which can be represented through content summary information, content hash values, or feature encoding information to achieve unique identification between different copyrighted works; and publication time information records the time when the work was first published.

[0057] Dynamic attributes are used to represent the real-time changes in the status of copyrighted works during the dissemination process, including initial dissemination data, real-time views, number of shares, number of favorites, number of comments, and public opinion rating. Initial dissemination data records the basic dissemination status of a copyrighted work when it enters the dissemination stage; real-time views, number of shares, and number of favorites reflect the current popularity of the copyrighted work; and the public opinion rating indicates the trend of user feedback on the copyrighted work. Dynamic attributes are updated in real-time as new data is acquired through the data collection interface, enabling the copyrighted work node to reflect changes in its dissemination status at different time stages.

[0058] A distribution platform node represents the platform entity that carries out the distribution of digital copyrights. Each distribution platform node corresponds to a distribution channel or medium. Distribution platform nodes are configured with basic attributes and dynamic attributes. Basic attributes include platform identification information, platform type information, and historical cooperation case information. Platform identification information is used to uniquely identify different distribution platforms; platform type information is used to distinguish short video platforms, long video platforms, text and image platforms, and other digital content distribution platforms; historical cooperation case information describes the historical cooperation between the distribution platform and copyrighted works.

[0059] The dynamic attributes of a dissemination platform node include platform user scale, real-time traffic data, dissemination rules, and traffic support policies. Platform user scale reflects the platform's overall influence; real-time traffic data reflects current platform traffic and content flow; dissemination rules represent dynamic factors influencing the dissemination of works, such as the platform's recommendation mechanism and content distribution rules; and traffic support policies represent the promotional measures taken by the platform for specific content. These dynamic attributes are updated based on data obtained from the dissemination platform's data interface to reflect changes in the dissemination platform environment.

[0060] User nodes represent user entities involved in digital copyright consumption, dissemination, and interaction. Each user node corresponds to a user object and is configured with basic and dynamic attributes. Basic attributes include user identification information, user level information, and spending power level information; dynamic attributes include user preference tags, historical copyright purchase records, and real-time browsing or interaction data.

[0061] Among them, user preference tags describe users' interest in different types of digital copyrighted content; historical copyright purchase records represent users' past copyright consumption behavior; and real-time browsing or interaction data reflect users' current level of interest in copyrighted works. By continuously updating the dynamic attributes of user nodes, a trajectory of user interest changes can be formed, providing dynamic user-side characteristics for subsequent copyright value assessment.

[0062] Legal entity nodes represent legal entities related to digital copyright, including copyright service agencies, licensors, rights protection entities, and other relevant legal entities. Legal entity nodes are configured with basic and dynamic attributes. Basic attributes include entity identification information, legal qualification certification number, and business scope information; dynamic attributes include policy interpretation capability score, infringement handling response timeliness, and the latest policy implementation details.

[0063] Among them, the policy interpretation capability score is used to represent the legal entity's ability to adapt to changes in copyright-related policies; the infringement handling response timeliness is used to represent the legal entity's efficiency in handling copyright disputes; and the latest policy implementation details are used to reflect changes in the current legal environment. By updating the dynamic attributes of legal entity nodes, the dynamic knowledge graph can be linked to the impact of policy environment changes on the value of digital copyrights.

[0064] After constructing the entity nodes, relationship edges are built based on the associated behaviors between different entity nodes to describe the connections between various entities in the digital copyright ecosystem. Relationship edges include dissemination relationship edges, consumption relationship edges, licensing relationship edges, and association relationship edges.

[0065] The propagation relationship edge represents the propagation relationship between the copyrighted work node and the propagation platform node. The propagation relationship edge points from the copyrighted work node to the propagation platform node. A corresponding propagation relationship edge is established when a copyrighted work is published, displayed, or propagated on a particular propagation platform. The propagation relationship edge is configured with relationship attributes, including the work's upload time, cumulative views, propagation duration, recommendation weight, and propagation status, used to record the propagation process of the copyrighted work on different propagation platforms.

[0066] Consumption relationship edges represent the consumption associations between user nodes and copyrighted work nodes, with each edge pointing from a user node to a copyrighted work node. A corresponding consumption relationship edge is established when a user engages in consumption activities such as watching, purchasing, subscribing to, or licensing a copyrighted work. These edges are configured with relationship attributes, including consumption timestamp, consumption type, consumption amount, consumption frequency, and user feedback information, which describe the user's actual consumption behavior of the copyrighted work.

[0067] Authorization relationship edges represent the authorization service relationship between copyrighted work nodes and legal entity nodes. An authorization relationship edge points from a copyrighted work node to a legal entity node. When a legal entity provides services such as rights confirmation, authorization management, infringement determination, or rights protection for a copyrighted work, a corresponding authorization relationship edge is established. Authorization relationship edges are configured with relationship attributes, including authorization effective time, authorization period, service type, and authorization status, used to record the legal service status corresponding to the copyrighted work.

[0068] Relationship edges represent the platform relationships between propagation platform nodes and user nodes, with each edge pointing from a propagation platform node to a user node. A corresponding relationship edge is established when a user accesses, interacts with, or consumes content on a propagation platform. These relationship edges are configured with relationship attributes, including user activity level, recent activity time, access frequency, and interaction intensity, which describe the degree of association between the user and the propagation platform.

[0069] To achieve unified management of digital copyright data and continuous recording of dynamic changes, a three-layer storage architecture of "node-edge-attribute" is adopted to store dynamic knowledge graphs, including node storage, edge storage, and attribute storage.

[0070] The node storage is used to store entity node data in the dynamic knowledge graph. Specifically, nodes are classified and managed according to entity type. Copyrighted work nodes, dissemination platform nodes, user nodes, and legal entity nodes are each assigned corresponding node tags, and each node is assigned a globally unique identifier as its primary key to ensure that entities from different data sources can be accurately linked.

[0071] Edge storage is used to store the relationships between entity nodes. Specifically, different types of relationship edges are defined according to propagation relationships, consumption relationships, authorization relationships, and association relationships. Each relationship edge is associated with at least a start node identifier, an end node identifier, and relationship type information, thereby forming a directional association network.

[0072] The attribute storage is used to store node attributes and relationship attributes. Node attributes are stored using a key-value pair structure, with basic and dynamic attributes stored in their respective attribute fields. Stable attributes such as work ID, platform type, and legal qualifications are stored in the basic attribute field; variable attributes such as real-time play counts, user preference tags, and platform traffic are stored in the dynamic attribute field. An attribute extension interface is reserved to support the addition of new feature dimensions in the future. Relationship attributes are also stored in key-value pairs to record play data in dissemination relationships, consumption amounts in consumption relationships, service status in licensing relationships, and user activity information in association relationships.

[0073] To ensure that the dynamic knowledge graph can continuously reflect changes in the digital copyright ecosystem, this embodiment also performs automatic update processing on the dynamic knowledge graph. The automatic update processing includes dynamic updates of node attributes, dynamic updates of relationship edges, and incremental updates of graph snapshots.

[0074] During the dynamic update of node attributes, newly added dynamic data is continuously acquired through a data acquisition interface, and the target node is determined based on the entity identifier corresponding to the data. The automatic update engine sends the cleaned and verified dynamic data to the graph database update interface, carrying the target node identifier and the attribute fields to be updated. The graph database locates the target node based on the node identifier and uses an incremental update method to modify only the dynamically changed attribute fields, while retaining the unchanged attribute fields. After the update is completed, update time information is recorded for the corresponding attribute data, and an attribute version record is generated to support subsequent historical status queries and time change analysis.

[0075] During the dynamic update of relationship edges, when a new entity association is detected, the automatic update engine generates a new relationship edge instruction based on business events. For example, when a copyrighted work is added to a distribution platform, a new distribution relationship edge is generated based on the work node identifier, platform node identifier, and distribution relationship type; when a user engages in new consumption behavior, a corresponding consumption relationship edge is generated. When an existing relationship becomes invalid, such as when a work is removed from a distribution platform or a user stops related activities, the automatic update engine generates a relationship invalidation instruction and marks the relationship edge status using logical deletion, rather than directly deleting historical relationship data, for use in subsequent time-series analysis.

[0076] In step S300, after completing the construction and continuous updating of the dynamic knowledge graph, in order to further describe the evolution of the factors influencing the value of digital copyright over time, this embodiment performs time-series processing on the dynamic knowledge graph based on a preset time window to generate several time-series graph snapshots.

[0077] Specifically, node data, relational edge data, and corresponding attribute data for the current time state are extracted from the dynamic knowledge graph according to a preset time window period (e.g., 6 hours). Entity nodes, relationships, and attribute states within the same time window are then combined to form a graph snapshot for an independent time state. Each time-series graph snapshot corresponds to the state of the digital copyright ecosystem at a certain time stage, and is used to record the relationships between copyrighted works, distribution platforms, users, and legal entities during that time stage.

[0078] The preset time window can be configured according to the rate of change of digital copyright data. For example, for short video copyright works with rapidly changing popularity, a shorter time window can be set to increase the frequency of snapshot generation; for long-term copyright works with stable dissemination status, a longer time window can be set to reduce data processing pressure.

[0079] In the specific generation process, the node status in the dynamic knowledge graph within the current time window is first obtained, including newly added nodes, deleted nodes, and node attribute changes; then the relationship edge status within the current time window is obtained, including newly added relationship edges, invalid relationship edges, and relationship attribute changes; finally, the above change information is associated with the graph status corresponding to the previous time window to generate a snapshot of the time series graph corresponding to the current time window.

[0080] To improve the efficiency of subsequent model processing, time series graph snapshots are stored using a unified data structure. Each graph snapshot includes at least a snapshot time identifier, a node set, a relation edge set, and an attribute feature set. The snapshot time identifier indicates the time period corresponding to the snapshot; the node set records the entity nodes existing within that time period; the relation edge set records the association states between nodes; and the attribute feature set records the dynamic changes in the nodes and relation edges.

[0081] By generating snapshots of the time-series graph from multiple consecutive time windows, a snapshot sequence is formed, enabling subsequent spatiotemporal graph neural network models to simultaneously acquire spatial correlation information and temporal evolution information during the digital rights dissemination process. Through this method, the originally static digital rights knowledge graph is transformed into a dynamic evolutionary graph with temporal continuity.

[0082] In step S400, after obtaining the time series map snapshot, in order to adapt to the dynamic changes of different digital copyright works in the dissemination process and avoid the problem of wasted computing resources or insufficient feature extraction accuracy caused by using a single fixed model, the number of dissemination platforms, the rate of change of core features, and the evaluation target label corresponding to the copyright work are first extracted based on the time series map snapshot. Based on the above scene features, the evaluation scene to which the current copyright work belongs is identified. Based on the identification results, a matching model is selected from spatiotemporal graph neural network models of different complexities for spatiotemporal feature learning to generate a spatiotemporal feature vector for digital copyright value assessment.

[0083] Specifically, the number of distribution platforms represents the coverage of a copyrighted work across different distribution channels. Statistical analysis is performed on the distribution relationship edges in the time-series graph snapshot to obtain the number of distribution platform nodes that establish distribution relationship edges with the target copyrighted work node. For example, when a copyrighted work has only one distribution platform node corresponding to a distribution relationship edge, it indicates that the copyrighted work is in a single-platform distribution state; when a copyrighted work establishes distribution relationship edges with multiple distribution platform nodes simultaneously, it indicates that the copyrighted work is in a multi-platform distribution state. The number of distribution platforms reflects the complexity of the copyrighted work's distribution network and serves as an important basis for judging the complexity of the model.

[0084] The core feature change rate is used to represent the degree of change of dynamic features related to digital copyright within a preset period. Specifically, dynamic attribute change information corresponding to copyrighted work nodes, dissemination platform nodes, and user nodes is extracted from multiple consecutive time series graph snapshots. Dynamic attributes include real-time play counts, forwarding counts, collection counts, public opinion ratings, user interaction frequency, and platform traffic changes. By comparing the changes of the same core feature within adjacent time windows, the corresponding feature change trend is determined. When the core feature changes significantly, it indicates that the copyrighted work may be in a stage of increasing popularity, dissemination, or market environment changes, requiring the use of a spatiotemporal graph neural network model with stronger dynamic feature extraction capabilities.

[0085] The assessment target tags are used to indicate the assessment purpose of the current digital copyright valuation task, including but not limited to existing copyright value accounting, real-time value tracking, and cross-platform commercial value assessment. The corresponding assessment target tags are obtained based on the assessment business needs, and combined with the number of dissemination platforms and the rate of change of core characteristics to determine the scenario type corresponding to the current copyrighted work.

[0086] In this embodiment, basic scene rules, core scene rules, and complex scene rules are pre-set, and the above scene rules are stored in the model selection configuration library.

[0087] The basic scenario rules are used to match copyright assessment scenarios with stable dissemination relationships and low dynamic changes. When the number of dissemination platforms corresponding to a copyrighted work is small, the change rate of core features within a preset period is low, and the assessment objective is stock value accounting, the current copyrighted work is identified as a basic assessment scenario. For example, if the copyrighted work has 1 dissemination platform; the change rate of core features in the past 7 days is less than 10%; and the assessment objective is stock value accounting, then it is determined to meet the basic scenario rules, and the ST-GCN model is selected.

[0088] The core scenario rules are used to match copyright assessment scenarios with obvious dynamic changes. When a copyrighted work is disseminated across multiple platforms, and the rate of change of core features is within a moderate range, while the assessment objective is real-time value tracking, the current copyrighted work is identified as part of the core assessment scenario. For example, if the copyrighted work is disseminated on 2-3 platforms; the rate of change of core features over the past 7 days is between 10% and 100%; and the assessment objective is real-time value tracking, then it is determined to meet the core scenario rules, and the AST-GNN model is selected.

[0089] Complex scenario rules are used to match copyright assessment scenarios with complex dissemination relationships and obvious long-term evolutionary characteristics. When a copyrighted work has multiple dissemination platform relationships, and simultaneously has IP derivative works, licensing chain extensions, or long-term commercial development relationships, and the core feature change rate is high, the current copyrighted work is judged as a complex assessment scenario. For example, if the following conditions are met: the number of dissemination platforms for the copyrighted work is greater than 3; there are IP derivative relationships; the core feature change rate in the past 7 days is greater than 100%; and the assessment objective is cross-platform commercial value assessment; then it is judged to meet the complex scenario rules, and the ST-TransGNN model is selected.

[0090] Furthermore, to avoid a decrease in model adaptability due to changes in copyright dissemination status after model selection, this embodiment also monitors the change rate of core features corresponding to copyrighted works and the number of dissemination platforms in real time, and triggers the switching of spatiotemporal graph neural network models according to preset scenario upgrade conditions.

[0091] Specifically, when a copyrighted work processed using the ST-GCN model experiences rapid changes in its core features—for example, a significant increase in metrics such as play counts and share counts within a preset period, causing the core feature change rate to exceed a preset upgrade threshold—it is determined that the current copyrighted work has been upgraded from a basic scenario to a dynamic scenario, and the model is switched from ST-GCN to AST-GNN. When a copyrighted work processed using the AST-GNN model adds multiple distribution platform relationships, causing the number of distribution platforms to exceed the complex scenario threshold, it is further switched to the ST-TransGNN model to improve the ability to extract cross-platform distribution relationships and long-term evolution patterns. For example, when a copyrighted work originally using the ST-GCN model experiences a play count change rate increase from 8% to 60%, exceeding the core scenario threshold, it automatically switches to the AST-GNN model; when a copyrighted work using the AST-GNN model adds new distribution platforms, increasing the number of distribution platforms from 2 to 4 and forming cross-platform distribution relationships, it automatically switches to the ST-TransGNN model.

[0092] During model switching, instead of rebuilding all the data, the model parameters corresponding to the target model are called based on the current time series map snapshot, and existing time series features are used for rapid initialization, so that the model switching process can adapt to the real-time assessment needs of digital copyright value.

[0093] When the ST-GCN model is selected based on scene recognition results, a short-time-series map snapshot sequence is chosen as the input data. This short-time-series map snapshot sequence includes map snapshots generated within multiple consecutive time windows, used to describe the changes in the dissemination status of copyrighted works over a short period.

[0094] Specifically, the input time-series graph snapshots are first preprocessed, including node feature normalization, missing data augmentation, and invalid snapshot filtering. For numerical dynamic attributes such as play counts, share counts, and favorite counts, normalization is used to eliminate the influence between different data units. For node attributes that do not change within certain time windows, they are augmented by associating node attribute information. Empty snapshots without node or relationship changes are filtered out to reduce the computational load of the model.

[0095] In the spatial feature learning stage, the ST-GCN model employs a graph convolutional message passing mechanism to model the spatial relationships between copyrighted work nodes and associated nodes. Specifically, taking the copyrighted work node as the central node, it obtains its directly associated neighboring nodes based on the relational edges in the dynamic knowledge graph, including dissemination platform nodes, user nodes, and legal entity nodes, and aggregates the attribute features of the neighboring nodes.

[0096] The ST-GCN model aggregates only the features of first-order neighbor nodes of copyrighted works, that is, it aggregates node information that has direct dissemination, consumption, and licensing relationships with copyrighted works, in order to avoid introducing invalid information interference caused by excessively distant related nodes. Through graph convolution operations, it fuses the attributes of the copyrighted work itself with the attributes of related entities, and outputs a spatial feature vector representing spatial relationships.

[0097] In the temporal feature learning phase, the ST-GCN model uses temporal convolution to process the sequence of map snapshots within a continuous time window. It extracts the trend features of dynamic attributes of copyrighted work nodes over time using temporal convolution kernels. For example, it captures trends in play count growth, user interaction changes, and dissemination range changes through temporal convolution.

[0098] Subsequently, the spatial and temporal feature vectors are concatenated and fused to form the spatiotemporal feature vector output by the ST-GCN model. This spatiotemporal feature vector represents the spatial correlation and temporal variation characteristics of copyrighted works during short-cycle dissemination.

[0099] Preferably, the snapshot sequence length is set to 10; each snapshot time window is 6 hours; therefore, the input data covers a time range of 60 hours. Simultaneously, the input data is filtered, retaining only data with stable propagation paths and no obvious surges in popularity. The model parameters are set as follows: number of spatial graph convolutional layers: 2; hidden layer dimension: 128; temporal convolution kernel size: 3; temporal convolution stride: 1; training batch size: 32; learning rate: 0.001. In the spatial feature learning stage: the GCN graph convolution message passing mechanism is adopted. The copyrighted work node is used as the central node, aggregating only 1-hop neighbor nodes. 1-hop neighbor nodes include: directly associated propagation platform nodes; directly consuming user nodes; and authorized legal entity nodes. Through graph convolution operations, the attributes of the copyrighted work itself and the attributes of directly associated entities are fused, outputting a 64-dimensional spatial feature vector. In the temporal feature learning stage: a one-dimensional temporal convolution is used to learn the trends in play count and user interaction changes over 10 consecutive time windows. Since the basic scene changes relatively little, the focus is on learning periodic change patterns rather than capturing sudden changes. Output a 64-dimensional temporal feature vector. Further, the 64-dimensional spatial feature vector and the 64-dimensional temporal feature vector are fused in a 7:3 ratio to generate a 128-dimensional spatiotemporal feature vector. Of this, the basic attributes of the work account for 40%; spatial interaction features account for 30%; and temporal trend features account for 30%.

[0100] When the AST-GNN model is selected based on the scene recognition results, the incremental time-series graph snapshot sequence is chosen as the input data. The incremental time-series graph snapshot sequence only records newly added nodes, newly added relationship edges, and attribute change information between adjacent time windows to reduce duplicate data input and improve the model's real-time processing capability.

[0101] In the spatial feature learning stage, the AST-GNN model uses a graph attention network message passing mechanism to dynamically calculate the importance of different neighbor nodes.

[0102] Specifically, attention weights are calculated based on the relationship type between copyrighted work nodes and neighboring nodes, node attribute status, and historical influence. This results in higher feature contribution rates for dissemination platform nodes, user nodes, and legal entity nodes that have a greater impact on copyright value, while reducing the contribution rate for nodes that have a smaller impact on value.

[0103] Through the aforementioned spatial attention mechanism, the AST-GNN model can distinguish the differences in the impact of different related entities on copyright value and output a spatial feature vector containing key related entity information.

[0104] In the temporal feature learning stage, the AST-GNN model employs causal convolution combined with a temporal attention mechanism to analyze data change trends at different time stages. The importance of different time nodes is determined based on the rate of change of core features. For time nodes experiencing rapid growth in viewership, significant changes in public opinion, or policy changes, the weight of the corresponding time features is increased; for time nodes with stable changes, the influence weight is decreased.

[0105] Subsequently, the AST-GNN model fuses spatial and temporal feature vectors through a multi-head attention fusion mechanism, and retains the original attribute information of the copyrighted work through residual connections, avoiding the loss of basic information during deep feature learning, and finally generating spatiotemporal feature vectors for dynamic copyright value assessment.

[0106] Preferably, the input snapshot sequence length is set to 10; each time window corresponds to 6 hours; therefore, the input data covers the copyright dissemination change process within a 60-hour range. The AST-GNN model adopts an incremental snapshot input method, only saving the data added or changed in the current time window relative to the previous time window. When the number of views for a copyrighted work was 1 million in the previous time window, and the number of views increases to 1.5 million in the current time window, only the data change corresponding to the 500,000 additional views is recorded, without having to repeatedly load the complete historical data. The AST-GNN model focuses on introducing the following dynamic features: real-time view change data; user interaction behavior change data; public opinion sentiment score change data; dissemination platform traffic change data; policy update data. The model parameters are configured as follows: the number of spatial attention heads is set to 8; the number of temporal attention heads is set to 4; the spatial feature dimension is set to 256; the temporal feature dimension is set to 256; the training batch size is set to 16; the initial learning rate is set to 0.0005; and the incremental training learning rate decay coefficient is set to 0.1.

[0107] During the spatial feature learning phase, the weights for high-traffic platform nodes and high-spending user nodes are set to a range of 0.7–0.9; while the weights for low-traffic platform nodes and ordinary user nodes are set to a range of 0.1–0.3. This dynamic weight adjustment allows the model to highlight key dissemination channels and core user groups. After spatial attention calculation, a 128-dimensional spatial feature vector is output.

[0108] During the time feature learning phase, when the rate of change in play count at a certain time point reaches 50% or more, that time point is identified as a peak in popularity; when a policy announcement leads to changes in the copyright licensing environment, the corresponding time point is identified as a policy-affected time point. For these key time points, the time attention weight is set to 0.8–0.9. For ordinary time points with stable play counts and minimal changes, the weight is reduced. This approach improves the model's sensitivity to sudden dissemination events and policy changes. After time attention calculation, a 128-dimensional time feature vector is output.

[0109] A 128-dimensional spatial feature vector and a 128-dimensional temporal feature vector are fused. Spatial attention employs an 8-head attention structure; temporal attention employs a 4-head attention structure. The fusion weights are dynamically adjusted based on the contribution of different features. Simultaneously, a residual connection structure is introduced to directly incorporate the original attribute features of the copyrighted work into the fusion process, ultimately outputting a 256-dimensional spatiotemporal feature vector. This 256-dimensional spatiotemporal feature vector simultaneously contains: the copyrighted work's own attribute information; interaction information from the dissemination platform; user behavior change information; and evolutionary information at key time nodes.

[0110] When the ST-TransGNN model is selected based on scene recognition results, a long-term time-series graph snapshot sequence is chosen as the input data. This long-term time-series graph snapshot sequence is used to describe the dissemination, platform migration, and evolution of commercial value of copyrighted works over a longer period.

[0111] Since there are many related nodes in complex scenarios, directly inputting all node information would cause a large computational burden. Therefore, this embodiment first uses a graph sampling aggregation algorithm to sample the nodes in the dynamic knowledge graph, starting from the copyrighted work node and selecting neighboring nodes with high relevance to form a subgraph structure for model calculation.

[0112] In the spatial feature learning stage, the ST-TransGNN model calculates the association weights between different nodes through a self-attention mechanism, capturing long-distance association features such as cross-platform propagation relationships, IP derivative relationships, and long-term business cooperation relationships.

[0113] In the temporal feature learning stage, the ST-TransGNN model uses a Transformer temporal encoder and a multi-head self-attention mechanism to analyze the correlation between different time windows and capture the trend changes in the long-term dissemination of copyrighted works.

[0114] Subsequently, through a cross-spatiotemporal self-attention fusion layer, spatial feature vectors and temporal feature vectors are jointly modeled, enabling the model to simultaneously consider changes in the propagation network structure and long-term temporal evolution, thereby generating high-dimensional spatiotemporal feature vectors for complex scenarios.

[0115] Preferably, the number of input snapshots is set to 30; each time window is 6 hours; and the overall coverage spans 180 hours of the dissemination process. Input data includes: cross-platform dissemination relationships; relationships between IP derivative works; data on the impact of long-term policy changes; and historical user consumption trends. Model parameters are configured as follows: 12 Transformer self-attention heads; 6 Transformer encoder layers; 1024 dimensions for the feedforward neural network; a training batch size of 8; and a learning rate of 0.0001.

[0116] For spatial long-distance correlation feature learning, the GraphSAGE graph sampling and aggregation algorithm is first employed. Specifically, the copyrighted work node is used as the central node; related nodes within a 2-hop range are sampled; the number of nodes sampled at each level is set to 10. The first-order neighbors include nodes of the direct propagation platform; the second-order neighbors include user groups and partner organization nodes associated with the propagation platform. Sampling reduces the number of nodes while preserving the main propagation paths. Subsequently, the long-distance correlation weights between nodes are calculated using the Transformer self-attention mechanism. The focus is on learning: cross-platform propagation linkage relationships; IP derivative value transmission relationships; and user group migration relationships. After spatial feature learning, a 256-dimensional spatial feature vector is output.

[0117] In the temporal feature learning phase, a Transformer temporal encoder is employed. The model can identify the long-term correlations between: early-stage growth in popularity; mid-stage expansion of user base; and late-stage increase in commercial licensing. A 12-head self-attention mechanism is used to model the evolutionary relationships between 30 time snapshots. The output is a 256-dimensional temporal feature vector.

[0118] A 256-dimensional spatial feature vector and a 256-dimensional temporal feature vector are input into a cross-spatiotemporal self-attention fusion layer. Through a Transformer fusion structure, the correlation between spatial propagation relationships and temporal evolution trends is calculated; the impact of cross-platform propagation on long-term value changes is learned; and information on the commercial value extension brought about by IP derivative relationships is integrated. Finally, a 512-dimensional spatiotemporal feature vector is generated. This 512-dimensional feature vector is used to describe the following in complex digital copyright scenarios: the impact of multi-platform propagation; long-term user value changes; commercial licensing diffusion paths; and the influence of policy environment trends.

[0119] To enable the spatiotemporal graph neural network model to adapt to the continuously changing digital copyright dissemination environment, this embodiment further sets up an incremental training mechanism to automatically trigger model parameter updates based on the addition of time series graph snapshots and changes in the core features of copyrighted works.

[0120] Specifically, the system continuously monitors the number of new time-series graph snapshots. When the number of new snapshots reaches a preset threshold, incremental training is triggered. For example, when the number of new snapshots reaches 10, it indicates that new propagation relationships, user behavior, and platform change data have been accumulated. The new time-series graph snapshots are then used as incremental training samples and input into the current spatiotemporal graph neural network model.

[0121] The system periodically monitors changes in key features of copyrighted works, including views, shares, user interactions, and public opinion ratings. The rate of change of these key features is calculated at preset intervals (e.g., 24 hours). When this rate exceeds a preset threshold, a significant change in the current copyright dissemination status is identified, triggering incremental training. For example, when the absolute value of the view change rate exceeds 50%, the model is updated to adapt to emerging trends, widespread dissemination, or changes in market trends.

[0122] During incremental training, snapshots of newly added time-series graphs are acquired, and parameters near the output layer of the spatiotemporal graph neural network model are adjusted, including the weights of the spatiotemporal feature fusion layer, attention weights, and output mapping parameters. Simultaneously, the feature extraction parameters of the lower-level nodes are frozen to avoid increased computational resource consumption due to full retraining. By fine-tuning local parameters, the model can quickly learn new copyright propagation patterns.

[0123] After completing incremental training, the updated spatiotemporal graph neural network model is used to re-extract the spatial correlation features and temporal evolution features from the latest graph snapshot, generate an updated spatiotemporal feature vector, and send the updated spatiotemporal feature vector to the copyright value assessment model, so that the copyright value assessment model can perform value calculation based on the latest dynamic data.

[0124] In step S500, ST-GCN outputs a 128-dimensional spatiotemporal feature vector; AST-GNN outputs a 256-dimensional spatiotemporal feature vector; and ST-TransGNN outputs a 512-dimensional spatiotemporal feature vector. Since the output dimensions of different models are different, feature dimension unification is first performed.

[0125] Specifically, a feature mapping layer maps spatiotemporal feature vectors of different dimensions to a unified feature space. Then, these spatiotemporal feature vectors are fused with fundamental copyright attribute features. These fused features include work type, publication time, dissemination scope, user spending power, licensing status, and legal risk status. A feature fusion layer combines these data to form comprehensive digital copyright value features.

[0126] Furthermore, the fused features are input into the copyright valuation model. The model predicts value based on these input features. The model output includes: a comprehensive digital copyright value score, a value level, and a value change trend. Specifically: the comprehensive digital copyright value score represents the current market value of the copyrighted work; the value level categorizes copyrights as high-value, medium-value, and low-value; and the value change trend predicts the direction of value growth or decline in the future.

[0127] The copyright value assessment model is used to quantitatively assess the comprehensive value of digital copyrighted works based on the spatiotemporal feature vectors output by the spatiotemporal graph neural network model. The copyright value assessment model can employ a value regression model based on a deep neural network, including an input layer, a feature fusion layer, a hidden layer, and an output layer. The input layer receives the spatiotemporal feature vectors output by the spatiotemporal graph neural network model, which include the copyrighted work's own attribute features, dissemination platform association features, user interaction features, and temporal evolution features. The feature fusion layer fuses feature information from different dimensions to obtain a comprehensive feature representation that characterizes the current market state of the copyrighted work. The hidden layer learns the correlation between the comprehensive feature representation and the copyright value based on a nonlinear mapping relationship. The output layer outputs the digital copyright value assessment result.

[0128] The copyright valuation model uses historical copyright transaction data, licensing price data, advertising cooperation price data, and dissemination effect data as training samples during the training phase. Specifically, the spatiotemporal feature vectors corresponding to historical copyrighted works are used as model input, while actual transaction value, licensing amount, or market valuation value are used as supervisory labels. The model parameters are trained through supervised learning, enabling the copyright valuation model to learn the mapping relationship between spatiotemporal features and copyright value.

[0129] During model training, historical copyright samples are divided into training, validation, and test sets according to a preset ratio, such as 70%, 15%, and 15%, respectively. The training set is used to iteratively optimize model parameters, the validation set is used to adjust model structure parameters, and the test set is used to verify the model's evaluation accuracy. The mean squared error loss function is used during model training to optimize the deviation between predicted and actual values. Model training is complete when the model's prediction error meets a preset threshold.

[0130] The input data format of the copyright valuation model is a fixed-dimensional feature vector format, such as a 256-dimensional AST-GNN spatiotemporal feature vector or a 512-dimensional ST-TransGNN spatiotemporal feature vector, and the corresponding input dimension is automatically matched according to different scenarios. The output data format is a digital copyright valuation result, including a comprehensive copyright value score, a predicted licensing price range, and a value change trend. The comprehensive copyright value score represents the current value level of the copyrighted work, the predicted licensing price range assists in copyright transaction pricing, and the value change trend reflects the growth or decline in copyright value as the dissemination status changes.

[0131] In a specific embodiment, the data processing method for the digital copyright assessment model proposed in this application is described using the real-time evaluation of the copyright value of a popular food short video on a short video platform as an application scenario. The evaluation object is a short video work published by a food creator on May 1, 2025, with the work number FP20250501. The video content is the process of making local specialty food. The objective is to conduct a real-time evaluation of the licensing transaction price and commercial cooperation value of the short video based on multi-dimensional dynamic factors such as the work's dissemination scope, user interaction, platform dissemination capabilities, and copyright service status.

[0132] In this embodiment, the system is deployed in a cloud computing environment, where the dynamic knowledge graph is stored using a Neo4j graph database cluster. The graph database server is configured with two cloud servers each with 24 CPU cores, 128GB of memory, and 2TB of storage, to support high-concurrency queries and dynamic updates of copyrighted work nodes, dissemination platform nodes, user nodes, and legal entity nodes. The spatiotemporal graph neural network model is deployed on a GPU training server, using two servers configured with NVIDIA A100 GPUs, each with 40GB of VRAM, for AST-GNN model training and subsequent incremental fine-tuning. The data acquisition server is configured with four cloud servers each with 16 CPU cores and 32GB of memory, using a multi-threaded acquisition program to obtain dissemination platform data, user behavior data, public opinion data, and policy data. The application server is configured with two cloud servers each with 8 CPU cores and 16GB of memory, used for model scheduling, data update triggering, and evaluation result output.

[0133] First, multi-source heterogeneous data related to the target copyrighted work is acquired through data collection interfaces. Specifically, copyrighted work data is obtained through a copyright management interface, including information such as work ID, work type, publication date, and content hash value; dissemination platform data is obtained through the platform's open API, including information such as play counts, share counts, favorite counts, platform user scale, and traffic support policies; user behavior data is obtained through a user behavior analysis interface, including user browsing history, interaction behavior, interest tags, and consumption behavior; and legal entity data is obtained through a copyright service platform, including the copyright agency's qualifications, authorization status, infringement handling capabilities, and the latest policy information.

[0134] Different collection cycles are configured based on the frequency of data changes for different types of data. Among them, core characteristic data such as real-time play counts, repost counts, and public opinion scores change rapidly, so the collection frequency is set to 1 hour / time; data with slower changes, such as platform user scale and copyright agency qualifications, are collected at a frequency of 24 hours / time; and data involving changes in copyright policies are collected at a frequency of 4 hours / time.

[0135] For example, at 10:00 AM on May 1, 2025, basic data for the work FP20250501 was collected for the first time. This data included the work's release time of 10:00 AM on May 1, 2025, initial views of 1000, shares of 50, favorites of 30, and a user opinion rating of 4.2. Simultaneously, data from the distribution platform was obtained, confirming that the work was first published on a specific content creation platform with approximately 600 million users, approximately 120 million active users at the current time, and a 10% traffic support ratio for food-related content.

[0136] A dynamic knowledge graph is constructed based on the collected data, and different types of data are converted into corresponding entity nodes.

[0137] First, create a copyrighted work node FP20250501. This node contains basic attributes such as work ID, work type, publication time, and hash value of work content features. It also configures dynamic attributes such as play count, forward count, collection count, and public opinion rating.

[0138] Next, create a node on a content creation and distribution platform. The basic attributes of the node include the platform ID and platform type, while the dynamic attributes include the current user base, real-time traffic status, and platform recommendation rules.

[0139] Simultaneously, user nodes were created based on user interaction data, selecting the top 1000 users who interacted with the work as associated user nodes. These user nodes included basic attributes such as user level and spending power level, as well as dynamic attributes such as food preference tags and browsing behavior records.

[0140] Furthermore, legal entity nodes are created, such as a copyright service company, to represent the relationship between copyright confirmation and licensing services. This node includes basic attributes such as copyright qualification number and business scope, as well as dynamic attributes such as policy interpretation ability score and infringement handling response time.

[0141] After the nodes are created, relationship edges are constructed based on the associations between entities. Specifically, this includes: establishing a dissemination relationship edge from the copyrighted work node FP20250501 to a creation platform node, representing the dissemination of the work on the creation platform; establishing a consumption relationship edge from the user node to the copyrighted work node, representing user viewing, collection, or interaction behaviors; establishing an authorization relationship edge from the copyrighted work node to the copyright service company, representing copyright confirmation services; and establishing an association relationship edge from a creation platform node to the user node, representing the user's active relationships on the platform.

[0142] Ultimately, a dynamic knowledge graph is formed, comprising four types of nodes: copyrighted works, dissemination platforms, users, and legal entities, as well as four types of edge relationships: dissemination, consumption, authorization, and association.

[0143] During the dissemination of the work, new dynamic data is continuously received, and the knowledge graph is incrementally updated through an automatic update engine.

[0144] For example, at 12:00 on May 1, 2025, the data collection interface detected that the number of views for the work FP20250501 increased from 1,000 to 100,000, the number of reposts increased to 8,000, and the public opinion rating improved to 4.8. This data change was sent to the Neo4j database, updating only the dynamic attribute fields corresponding to the copyrighted work node, and recording the update time.

[0145] Subsequently, at 15:00 on May 1, 2025, it was detected that the work had been added to the Kuaishou distribution channel. Therefore, a new distribution relationship edge was automatically created, connecting the copyright work node FP20250501 to the Kuaishou platform node, and writing relationship attributes such as the new distribution time and initial play count.

[0146] Time series graph snapshots are generated according to a 6-hour time window. For example, a third graph snapshot is generated at 16:00 on May 1, 2025. This snapshot does not repeatedly store the complete graph, but only records the differences in data such as the Kuaishou dissemination relationship, changes in play volume, and changes in user interaction compared to the previous time window.

[0147] Extract the number of propagation platforms, the rate of change of core features, and the target labels for evaluation from time series map snapshots.

[0148] In this embodiment, the work FP20250501 was initially only disseminated on a certain creation platform. However, due to the rapid increase in the number of views from 1,000 to 100,000, the core feature change rate reached 990%. At the same time, the evaluation target was real-time licensing pricing. Therefore, it was determined that the work belonged to a high-dynamic core evaluation scenario.

[0149] According to the preset scenario rules, works that are disseminated on a single platform and have a low rate of change are matched with the ST-GCN model; works that are dynamically disseminated on multiple platforms are matched with the AST-GNN model; and works that are disseminated across platforms and involve IP derivative value assessment are matched with the ST-TransGNN model.

[0150] Since the current work meets the core scene conditions, the AST-GNN model is automatically selected for spatiotemporal feature learning.

[0151] The most recent 10 time series map snapshots are used as input to the AST-GNN model, covering a 60-hour propagation and change process. Numerical features such as play count and forward count are normalized, and missing data are supplemented by the mean of neighboring nodes.

[0152] The AST-GNN model parameters were set to 8 spatial attention heads, 4 temporal attention heads, 256 spatial and temporal feature dimensions, 16 training batches, and a learning rate of 0.0005.

[0153] During spatial feature learning, the model employs a graph attention network mechanism to calculate the importance of copyrighted work nodes and related entity nodes. For example, due to the high dissemination capabilities of a certain creation platform, a node of that platform is assigned an attention weight of 0.8; nodes of highly interactive users are assigned a weight of 0.7, enabling the model to focus on entities with high dissemination influence.

[0154] During the temporal feature learning process, the model employs a TCN causal convolution combined with a temporal attention mechanism to focus on analyzing the rapid growth phase of play counts. For example, it detects an explosive increase in play counts between 12:00 and 15:00, and therefore assigns a temporal attention weight of 0.9 to the data during this time period.

[0155] By fusing spatial attention and temporal attention, a 256-dimensional spatiotemporal feature vector is generated, which includes features of the work's dissemination range, user interaction, platform influence, and popularity trend.

[0156] As the work continues to spread, we will continue to monitor changes in the map.

[0157] At 10:00 AM on May 2, 2025, a new Bilibili distribution channel was detected for the work, bringing the total number of distribution platforms to three, and the core feature change rate reached 1500%. Based on the complex scenario rules, the work was upgraded from the core evaluation scenario to the complex evaluation scenario.

[0158] Therefore, the model is automatically switched from AST-GNN to ST-TransGNN, and the ST-TransGNN pre-trained parameters are invoked. Subsequently, the newly added cross-platform propagated data is input into the model to perform incremental training.

[0159] During incremental training, the feature extraction parameters of the bottom-level nodes are frozen, and only the parameters of the Transformer attention layer and spatiotemporal fusion layer are adjusted, enabling the model to quickly learn cross-platform propagation rules. The entire incremental training process is kept under one hour.

[0160] After completing the spatiotemporal feature extraction, the 512-dimensional spatiotemporal feature vector output by ST-TransGNN is input into the copyright value assessment model.

[0161] The copyright value assessment model comprehensively analyzes factors such as cross-platform dissemination capabilities, user interest matching degree, public opinion trends, copyright policy environment, and licensing service capabilities in the vector, and outputs real-time copyright value assessment results.

[0162] Ultimately, for work FP20250501, its current licensing transaction value is assessed to be between RMB 15,000 and RMB 20,000, and its commercial advertising placement cooperation value is approximately RMB 30,000 per piece.

[0163] Based on another preferred embodiment described above, see [link to preferred embodiment]. Figure 2 As shown, this embodiment provides a data processing system for a digital rights management (DRM) assessment model, used to apply the above-described data processing method for a DRM assessment model, including:

[0164] The data acquisition module is configured to acquire multi-source heterogeneous data, including copyrighted work data, distribution platform data, user behavior data, and legal entity data.

[0165] The dynamic knowledge graph construction module is configured to transform copyrighted work data, dissemination platform data, user behavior data, and legal entity data into entity nodes, and construct relationship edges based on the dissemination relationship, consumption relationship, authorization relationship, and association relationship between each entity node to form a dynamic knowledge graph;

[0166] The snapshot generation module is configured to perform time-series processing on the dynamic knowledge graph based on a preset time window, generating several time-series graph snapshots.

[0167] The model adaptive selection module is configured to extract the number of propagation platforms, the rate of change of core features, and the evaluation target label based on the time series graph snapshot, and select a spatiotemporal graph neural network model; and input the time series graph snapshot into the selected spatiotemporal graph neural network model to perform spatiotemporal feature learning and generate spatiotemporal feature vectors.

[0168] The digital copyright assessment module is configured to perform feature fusion processing on spatiotemporal feature vectors and input them into the copyright value assessment model to obtain the digital copyright value assessment results.

[0169] In summary, the dynamic knowledge graph integrates heterogeneous entities from multiple sources, including copyrighted works, dissemination platforms, users, and legal entities, breaking through the limitations of traditional assessments that only focus on the work itself and capturing the interactive influence of related entities. ST-GNN achieves joint learning of spatial structural relationships and temporal series features, and the generated spatiotemporal feature representation vectors better reflect the dynamic changes in copyright value, demonstrating stronger representation capabilities compared to single static or temporal features. The automatic update engine automates the entire process of data collection, graph updates, and model evolution without manual intervention, ensuring that feature data and model representations are always up-to-date and adaptable to dynamic changes such as market trends and policy adjustments. Comprehensive, dynamic, and in-depth feature representations provide higher-quality input for subsequent assessment models, improving the assessment model's adaptability to complex dynamic scenarios and its assessment accuracy.

[0170] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the present invention.

Claims

1. A data processing method for a digital rights management (DRM) assessment model, characterized in that, include: Acquire multi-source heterogeneous data; the multi-source heterogeneous data includes copyrighted work data, dissemination platform data, user behavior data, and legal entity data; Copyrighted work data, dissemination platform data, user behavior data, and legal entity data are transformed into entity nodes, and relationship edges are constructed based on the dissemination relationship, consumption relationship, authorization relationship, and association relationship between the entity nodes to form a dynamic knowledge graph; The dynamic knowledge graph is processed in a time sequence based on a preset time window to generate several time-series graph snapshots. The number of propagation platforms, the rate of change of core features, and the target labels are extracted based on the time series graph snapshot, and a spatiotemporal graph neural network model is selected. The time series graph snapshot is then input into the selected spatiotemporal graph neural network model to learn spatiotemporal features and generate spatiotemporal feature vectors. The spatiotemporal feature vectors are subjected to feature fusion processing and input into the copyright value assessment model to obtain the digital copyright value assessment result; When selecting a spatiotemporal graph neural network model, the following should be included: The number of dissemination platforms for copyrighted works, the rate of change of core features within a preset period, and the evaluation target label are extracted from the time series graph snapshot. The number of dissemination platforms, the rate of change of core features, and the evaluation target label are compared with the basic scene rules, the core scene rules, and the complex scene rules to determine the scene type. Based on the determined scene type, the matching spatiotemporal graph neural network model is selected from the ST-GCN model, the AST-GNN model, and the ST-TransGNN model. The system monitors the rate of change of the core features and the number of propagation platforms in real time, and triggers the adaptive switching of the spatiotemporal graph neural network model when the scenario upgrade conditions are met.

2. The data processing method for a digital rights assessment model according to claim 1, characterized in that, The entity nodes include copyrighted work nodes, dissemination platform nodes, user nodes, and legal entity nodes, and each entity node is configured with basic attributes and dynamic attributes; The relationship edges include dissemination relationship edges, consumption relationship edges, licensing relationship edges, and association relationship edges; the dissemination relationship edge points from the copyrighted work node to the dissemination platform node; the consumption relationship edge points from the user node to the copyrighted work node; the licensing relationship edge points from the copyrighted work node to the legal entity node; and the association relationship edge points from the dissemination platform node to the user node.

3. The data processing method for a digital rights assessment model according to claim 1, characterized in that, The dynamic knowledge graph adopts a three-layer storage architecture of node-edge-attribute, including node storage, edge storage, and attribute storage.

4. The data processing method for a digital rights assessment model according to claim 1, characterized in that, The dynamic knowledge graph is dynamically updated, including dynamic updates of node attributes, dynamic updates of edge relationships, and incremental generation and updating of graph snapshots.

5. The data processing method for a digital rights assessment model according to claim 1, characterized in that, When generating the spatiotemporal feature vector, the process includes: when the selected spatiotemporal graph neural network model is the ST-GCN model, selecting a short-time snapshot sequence as input data; in the spatial feature learning stage, using a graph convolution message passing mechanism, with the copyrighted work node as the central node, aggregating the features of 1-hop neighbor nodes, and outputting a spatial feature vector; in the temporal feature learning stage, capturing the temporal patterns within a short period of the copyrighted work through one-dimensional convolution, and outputting a temporal feature vector; and concatenating the spatial feature vector and the temporal feature vector at a fixed ratio to generate the first-dimensional spatiotemporal feature vector.

6. The data processing method for a digital rights management model according to claim 1, characterized in that, When generating the spatiotemporal feature vector, the method further includes: when the selected spatiotemporal graph neural network model is an AST-GNN model, selecting an incremental temporal snapshot sequence as input data; in the spatial feature learning stage, using a graph attention network message passing mechanism, calculating the attention coefficients between the copyrighted work node and its neighboring nodes and assigning weights, and outputting the spatial feature vector; in the temporal feature learning stage, using a causal convolution combined with a temporal attention mechanism, assigning weights based on the rate of change of the core features, and outputting the temporal feature vector; and using multi-head attention weighted fusion and residual connection to fuse the spatial feature vector and the temporal feature vector to generate the second-dimensional spatiotemporal feature vector.

7. The data processing method for a digital rights assessment model according to claim 1, characterized in that, When generating the spatiotemporal feature vector, the method further includes: when the selected spatiotemporal graph neural network model is the ST-TransGNN model, selecting a long-term snapshot sequence as input data, and using a graph sampling aggregation algorithm to sample and aggregate node features; in the spatial feature learning stage, calculating the long-distance correlation weights between nodes through a self-attention mechanism to capture cross-platform propagation linkage features and outputting a spatial feature vector; in the temporal feature learning stage, using a Transformer temporal encoder to capture the feature evolution law within a long period through a self-attention mechanism and outputting a temporal feature vector; and using a cross-spatiotemporal self-attention fusion layer to fuse the spatial feature vector and the temporal feature vector to generate the third-dimensional spatiotemporal feature vector.

8. The data processing method for a digital rights assessment model according to claim 1, characterized in that, This also includes incremental training of the spatiotemporal graph neural network model: Incremental training is triggered when the number of newly added snapshots in the time series map exceeds a certain threshold. Incremental training is triggered when the absolute value of the rate of change of the core features of key nodes within a preset period exceeds the rate of change threshold. Based on the newly added time series graph snapshot, the parameters of the spatiotemporal graph neural network model near the output layer are adjusted, and the parameters of the spatiotemporal graph neural network model are updated.

9. A data processing system for a digital rights management (DRM) assessment model, used to apply the data processing method for a DRM assessment model as described in any one of claims 1-8, characterized in that, include: The data acquisition module is configured to acquire multi-source heterogeneous data, including copyrighted work data, dissemination platform data, user behavior data, and legal entity data. The dynamic knowledge graph construction module is configured to transform copyrighted work data, dissemination platform data, user behavior data, and legal entity data into entity nodes, and construct relationship edges based on the dissemination relationship, consumption relationship, authorization relationship, and association relationship between the entity nodes to form a dynamic knowledge graph. The snapshot generation module is configured to perform time-series processing on the dynamic knowledge graph based on a preset time window to generate several time-series graph snapshots. The model adaptive selection module is configured to extract the number of propagation platforms, the rate of change of core features, and the evaluation target label based on the time series graph snapshot, and select a spatiotemporal graph neural network model; and input the time series graph snapshot into the selected spatiotemporal graph neural network model to perform spatiotemporal feature learning and generate spatiotemporal feature vectors. The digital copyright assessment module is configured to perform feature fusion processing on the spatiotemporal feature vector and input it into the copyright value assessment model to obtain the digital copyright value assessment result.

Images