A hyperbolic diffusion network-based time link prediction method and device

Abstract

This invention discloses a method and apparatus for predicting temporal links based on hyperbolic diffusion networks. The method includes acquiring a temporal topology graph corresponding to a historical time period and dividing it into multiple temporal stages; locating graph nodes exhibiting forwarding behavior in each temporal stage from the temporal topology graph, extracting the cascaded topology graph and data forwarding subgraph corresponding to the graph nodes, optimizing the Euclidean eigenvectors corresponding to each graph node to generate a global hyperbolic embedding matrix; extracting the node stage features corresponding to the currently active graph nodes from the global hyperbolic embedding matrix; updating the node stage features by perturbation according to a preset fluctuation perturbation operator to obtain a data prediction feature vector; calculating the similarity between the data prediction feature vector and the candidate global hyperbolic embedding matrix, and locating target predicted graph nodes with a similarity greater than or equal to a preset similarity threshold, thereby accurately reflecting the hierarchical transitions of graph data propagation links and improving resource allocation and utilization.

Description

Technical Field

[0001] This invention relates to the field of graph data processing technology, and in particular to a method and apparatus for predicting time-series links based on hyperbolic diffusion networks. Background Technology

[0002] With the advent of the big data era, the amount of data generated and recorded in human production and life has exploded. Due to the ability of graph data as a data structure to represent entities and relationships, it is possible to abstract the data of complex systems in the real world into graph data for analysis. It is widely used in fields such as transportation systems, social networks, e-commerce, and transaction data analysis.

[0003] However, in practice, the topology of graph data is always in a dynamic evolution process, exhibiting high dynamism and adversarial nature. Current sequence prediction models are often only applicable to Euclidean space data, and their model structures are mostly transductive structures that rely on fixed node IDs to learn embedding vectors.

[0004] In scale-free graph data with deep hierarchical structures, traditional Euclidean space embedding is difficult to maintain the hierarchical structure losslessly in low-dimensional space. This leads to confusion between nodes at different levels in the embedding space, and it is difficult to generate effective representations when faced with one-off nodes, thus failing to accurately reflect the hierarchical transitions in the graph data propagation chain. Summary of the Invention

[0005] This invention provides a temporal link prediction method and apparatus based on hyperbolic diffusion networks, which solves the technical problem that traditional Euclidean space embedding is difficult to maintain the hierarchical structure losslessly in low-dimensional space in graph data with scale-free characteristics and deep hierarchical structure. This leads to confusion of nodes at different levels in the embedding space, and it is difficult to generate effective representations when faced with one-off nodes, thus failing to accurately reflect the hierarchical transitions of graph data propagation links.

[0006] This invention provides a time-series link prediction method based on hyperbolic diffusion networks, comprising:

[0007] Obtain the time series topology graph corresponding to the historical time period and divide it into multiple time series stages;

[0008] From the time-series topology graph, locate the graph nodes that exhibit forwarding behavior in each of the time-series stages, and extract the cascaded topology graph and data forwarding subgraph corresponding to the graph nodes;

[0009] Based on the cascaded topology graph and the data forwarding subgraph, the Euclidean eigenvectors corresponding to each graph node are optimized to generate a global hyperbolic embedding matrix;

[0010] Extract the node phase features corresponding to the currently active graph node from the global hyperbolic embedding matrix;

[0011] The node's phased features are perturbed and updated according to a preset fluctuation perturbation operator to obtain a data prediction feature vector;

[0012] Calculate the similarity between the data prediction feature vector and the candidate global hyperbolic embedding matrix, and locate the target prediction graph node whose similarity is greater than or equal to a preset similarity threshold.

[0013] Optionally, the step of locating the graph nodes that exhibit forwarding behavior in each of the time-series stages from the time-series topology graph, and extracting the cascaded topology graph and data forwarding subgraph corresponding to the graph nodes, includes:

[0014] Locate the graph nodes in the time sequence topology graph that exhibit forwarding behavior at each of the time sequence stages;

[0015] Extract the node hierarchy attributes corresponding to each graph node, and associate each graph node to construct a cascaded topology graph;

[0016] Based on the forwarding behavior attributes between each graph node and its adjacent graph nodes, a data forwarding subgraph corresponding to each graph node is constructed.

[0017] Optionally, the step of optimizing the Euclidean eigenvectors corresponding to each graph node based on the cascaded topology graph and the data forwarding subgraph to generate a global hyperbolic embedding matrix includes:

[0018] Based on the cascaded topology graph and the data forwarding subgraph, create a Euclidean feature vector corresponding to each graph node in Euclidean space;

[0019] The exponential mapping operator is used to project each of the Euclidean eigenvectors from the Euclidean space to the hyperbolic space to generate node hyperbolic embedding vectors.

[0020] An initial global hyperbolic embedding matrix is ​​constructed using all the hyperbolic embedding vectors of the nodes;

[0021] Using the cascaded topology graph and the data forwarding subgraph as constraints, the distance between the hyperbolic embedding vectors of each node in the initial global hyperbolic embedding matrix is ​​adjusted according to a preset hyperbolic distance loss function to generate a global hyperbolic embedding matrix.

[0022] Optionally, it also includes:

[0023] When a new graph node is added, the forwarding behavior attribute associated with the new graph node and the node level attribute it belongs to are extracted, and a new Euclidean feature vector corresponding to the new graph node is created.

[0024] The newly added Euclidean eigenvectors are projected onto the hyperbolic space using the exponential mapping operator to generate hyperbolic embedding vectors for the new nodes.

[0025] The node hyperbolic embedding vector is added to the global hyperbolic embedding matrix to obtain a new global hyperbolic embedding matrix.

[0026] Optionally, the node hyperbolic embedding vector is:

[0027] ;

[0028] in, Let be the hyperbolic embedding vector of graph node u, and K be a preset manifold curvature parameter. Let be the Euclidean eigenvector of graph node u. This is an exponential mapping operator.

[0029] Optionally, the step of extracting the node phase features corresponding to the currently active graph node from the global hyperbolic embedding matrix includes:

[0030] Extract the initial node features corresponding to the currently active graph node from the global hyperbolic embedding matrix;

[0031] The initial node features are averaged and pooled to obtain the node stage features.

[0032] Optionally, the step of updating the node's phased features by perturbation according to a preset fluctuation perturbation operator to obtain the data prediction feature vector includes:

[0033] The preset recurrent neural network is invoked to generate the initial data representation corresponding to the current time step based on the node's stage features and historical hidden state;

[0034] The initial data representation is weighted and fused according to dynamically adjusted weights and preset fluctuation disturbance operators to generate the intermediate data representation corresponding to the current time step;

[0035] A nonlinear mapping function is used to nonlinearly fuse the intermediate data representation with the expected estimate of the fluctuation perturbation operator to obtain the data prediction feature vector.

[0036] Optionally, after performing the step of weighting and fusing the initial data representation according to dynamically adjusted weights and a preset fluctuation disturbance operator to generate the intermediate data representation corresponding to the current time step, the method further includes:

[0037] The activation function is invoked to update the hidden state corresponding to the current time step as the new historical hidden state, based on the intermediate data representation and the weight matrix corresponding to the current time step.

[0038] Optionally, the candidate global hyperbolic embedding matrix is ​​composed of hyperbolic embedding vectors of multiple candidate graph nodes; the step of calculating the similarity between the data prediction feature vector and the candidate global hyperbolic embedding matrix, and locating the target prediction graph node whose similarity is greater than or equal to a preset similarity threshold, includes:

[0039] The inner product between the predicted feature vector of the data and the candidate global hyperbolic embedding matrix is ​​calculated as the similarity of each candidate graph node;

[0040] Multiple candidate graph nodes with a similarity greater than or equal to a preset similarity threshold are identified as intermediate graph nodes;

[0041] Based on the similarity, a preset number of intermediate graph nodes are selected as target prediction graph nodes, arranged from high to low.

[0042] The present invention also provides a time-series link prediction device based on hyperbolic diffusion networks, comprising:

[0043] The time series partitioning module is used to obtain the time series topology map corresponding to the historical time period and divide it into multiple time series stages;

[0044] The location extraction module is used to locate the graph nodes that exhibit forwarding behavior in each of the time-series stages from the time-series topology graph, and extract the cascaded topology graph and data forwarding subgraph corresponding to the graph nodes;

[0045] The global hyperbolic embedding matrix generation module is used to optimize the Euclidean eigenvectors corresponding to each graph node based on the cascaded topology graph and the data forwarding subgraph to generate a global hyperbolic embedding matrix.

[0046] The node stage feature extraction module is used to extract the node stage features corresponding to the currently active graph nodes from the global hyperbolic embedding matrix.

[0047] The data prediction feature vector generation module is used to update the stage features of the node according to the preset fluctuation perturbation operator to obtain the data prediction feature vector.

[0048] The graph node prediction module is used to calculate the similarity between the data prediction feature vector and the candidate global hyperbolic embedding matrix, and to locate the target predicted graph node whose similarity is greater than or equal to a preset similarity threshold.

[0049] As can be seen from the above technical solutions, the present invention has the following advantages:

[0050] This invention acquires a temporal topology graph corresponding to a historical time period and divides it into multiple temporal stages. It locates graph nodes exhibiting forwarding behavior in each temporal stage from the temporal topology graph, extracts the cascaded topology graph and data forwarding subgraph corresponding to each graph node, optimizes the Euclidean eigenvector corresponding to each graph node, and generates a global hyperbolic embedding matrix. It extracts the node stage features corresponding to the currently active graph nodes from the global hyperbolic embedding matrix, updates the node stage features by perturbation according to a preset fluctuation perturbation operator, and obtains a data prediction feature vector. It calculates the similarity between the data prediction feature vector and the candidate global hyperbolic embedding matrix, locates target predicted graph nodes with similarity greater than or equal to a preset similarity threshold, and thus, through the introduction of hyperbolic space and diffusion network, accurately reflects the hierarchical transitions of graph data propagation links while effectively predicting graph nodes for data forwarding. This enables pre-allocation of resources and improves resource utilization. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 A flowchart illustrating the steps of a time-series link prediction method based on hyperbolic diffusion networks, provided in an embodiment of the present invention;

[0053] Figure 2 A timing topology diagram provided in an embodiment of the present invention;

[0054] Figure 3 This is a schematic diagram of the hyperbolic space embedding and diffusion learning process in an embodiment of the present invention;

[0055] Figure 4 This is a structural block diagram of a time-series link prediction device based on a hyperbolic diffusion network, provided in an embodiment of the present invention. Detailed Implementation

[0056] This invention provides a temporal link prediction method and apparatus based on hyperbolic diffusion networks to address the technical problem that traditional Euclidean space embedding is difficult to maintain the hierarchical structure losslessly in low-dimensional space in graph data with scale-free characteristics and deep hierarchical structure. This leads to confusion between nodes at different levels in the embedding space, and it is difficult to generate effective representations when faced with one-off nodes, thus failing to accurately reflect the hierarchical transitions of graph data propagation links.

[0057] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0058] Please see Figure 1 and Figure 2 , Figure 1 This is a flowchart illustrating the steps of a time-series link prediction method based on hyperbolic diffusion networks, as provided in an embodiment of the present invention. Figure 2 This is a timing topology diagram provided in an embodiment of the present invention.

[0059] This invention provides a time-series link prediction method based on hyperbolic diffusion networks, comprising:

[0060] Step 101: Obtain the time series topology map corresponding to the historical time period and divide it into multiple time series stages;

[0061] A temporal topology graph refers to a dynamic graph structure constructed by integrating topological structure and temporal attributes, with graph nodes as vertices and data forwarding links as directed edges. It represents the connection relationships between graph nodes and the temporal characteristics of data forwarding within a historical period, such as... Figure 2 As shown, it includes initial graph nodes, intermediate graph nodes, and final graph nodes. Black graph nodes represent graph nodes that exhibit data forwarding behavior, while green graph nodes represent graph nodes that do not exhibit data forwarding behavior. Graph nodes can include, but are not limited to, communication network nodes, data forwarding nodes, blockchain nodes, or account addresses. The data forwarding link can be a timestamped traffic forwarding link, a data forwarding link, a blockchain data link, or a financial data flow link, etc.

[0062] A time series phase refers to at least two consecutive time units obtained by dividing the time dimension of the time series topology graph according to a fixed time window.

[0063] In this embodiment, by collecting node deployment information, traffic forwarding logs, link connectivity data, etc. from communication networks or blockchains during historical periods, a time-series topology graph containing graph nodes, link connection relationships, and time-series markers is constructed. Based on the time-series evolution of network traffic and modeling requirements, the time dimension of the time-series topology graph is regularly segmented to obtain multiple non-overlapping time-series stages that continuously cover historical periods.

[0064] In addition, the stage boundaries can be automatically identified based on the traffic mutation characteristics and node activity change characteristics of the time-series topology graph, so as to divide the time-series topology graph in an adaptive time-series stage.

[0065] In another example, the time-series topology can also consist of account addresses on the blockchain and related historical financial data flow records. This is achieved by dividing the flow of each piece of financial data into n stages according to a time window. The set of account addresses that newly received financial data during this period is represented as Ultimately, for specific financial data streams... The sampling results are expressed as The time-series topology diagram for traffic data streams is similar. Figure 2 As shown.

[0066] Step 102: Locate the graph nodes that exhibit forwarding behavior at each time stage from the time-series topology graph, and extract the cascaded topology graph and data forwarding subgraph corresponding to the graph nodes;

[0067] Forwarding behavior refers to the network operation in which a graph node receives data from an upstream node and transmits it to a downstream node.

[0068] Graph nodes refer to vertex entities in a time-series topology graph, corresponding to traffic forwarding entities such as network elements, terminals, and switches in a communication network, or entities such as account addresses in a financial data network.

[0069] A cascaded topology graph refers to a hierarchical association topology formed by hop-by-hop forwarding of data, representing the complete diffusion path structure of data from the source node to the end node.

[0070] A data forwarding subgraph refers to a local topology formed by extracting the direct upstream and downstream connections and forwarding links of a single target graph node, representing the direct data interaction characteristics of the nodes.

[0071] In this embodiment, by traversing the time-series topology graph data within each time-series stage, and based on the time-series markers and traffic transmission records of the links, the graph nodes that perform traffic reception and transmission operations within each time-series stage are located. Then, with the located graph nodes as the core, the complete hierarchical association structure formed by hop-by-hop traffic transmission is extracted from the time-series topology graph to obtain a cascaded topology graph. At the same time, the direct upstream and downstream connections and forwarding links of the nodes are extracted to form a local topology to obtain a data forwarding subgraph. This separates the global topology and local topology features, preserving the hierarchical transmission law of traffic diffusion while focusing on the direct forwarding attributes of the nodes, providing a hierarchical topology basis for subsequent feature optimization.

[0072] In addition, during the extraction of the data forwarding subgraph, redundant connections and isolated nodes without traffic interaction can be removed, simplifying the structure of the cascaded topology graph and the data forwarding subgraph.

[0073] In another example, the account addresses that send and receive financial data can be located at each time series stage as graph nodes. The cascaded topology graph formed by the step-by-step transfer of financial data can be extracted from the time series topology graph. At the same time, the local topology of each graph node and its upstream and downstream connections can be extracted to construct a data forwarding subgraph.

[0074] In one example of the present invention, step 102 may include the following sub-steps:

[0075] Locate the graph nodes in the time sequence topology graph that exhibit forwarding behavior at each time stage;

[0076] Extract the node hierarchy attributes corresponding to each graph node, and associate the graph nodes to construct a cascaded topology graph;

[0077] Construct a data forwarding subgraph for each graph node based on the forwarding behavior attributes between each graph node and its adjacent graph nodes.

[0078] Node hierarchy attributes refer to the characteristic attributes that represent the topological hierarchy, upstream and downstream relationships, and flow depth of a graph node in the data integrity forwarding link.

[0079] Forwarding behavior attributes refer to the core features that characterize the data forwarding interaction between adjacent graph nodes, including the connectivity of forwarding links, the direction of data flow, and the timing of interaction.

[0080] In this embodiment, under the scenario of traffic forwarding in a communication network, the full data of the time-series topology graph corresponding to each time-series stage is traversed. Based on the link time-series markers and traffic forwarding records, all graph nodes performing traffic reception and downstream forwarding operations within each time-series stage are accurately located. Simultaneously, the node-level attributes corresponding to each located graph node are extracted, including the network layer to which the node belongs, the number of traffic hops, and the upstream and downstream topology dependencies. Then, based on the time-series traffic forwarding associations between nodes, graph nodes with direct or indirect traffic transmission associations are cross-time-series associated, constructing a complete cascaded topology graph representing the hop-by-hop transmission pattern of traffic. Finally, according to... For each graph node, the forwarding behavior attributes such as traffic forwarding direction, link connectivity, and interaction timing between the node and its adjacent upstream and downstream nodes are used to construct a corresponding local data forwarding subgraph for each located graph node. By first locating active forwarding nodes, then extracting hierarchical attributes to construct the global topology, and finally generating local subgraphs, the step-by-step execution logic avoids invalid calculations for all silent nodes with no forwarding behavior, significantly reducing the computational overhead of topology construction. At the same time, through the hierarchical construction of global cascaded topology and local subgraphs, the global link rules of traffic transmission across levels are fully preserved, while the direct traffic interaction characteristics of a single node are accurately focused.

[0081] In another example, in a financial data forwarding scenario, the entire temporal topology graph corresponding to each time series stage is traversed. Based on transaction time series markers and financial data flow records, all graph nodes performing financial data reception and downstream forwarding operations within each time series stage are accurately located. Simultaneously, the node hierarchy attributes corresponding to each located graph node are extracted, including the depth of financial data flow, the topological hierarchy of the account in the transaction chain, and the upstream and downstream financial data subordination relationships. Then, based on the temporal transaction forwarding associations between nodes, graph nodes with direct or indirect financial data transmission associations are cross-time series-associated to construct a complete cascaded topology graph representing the hop-by-hop flow pattern of financial data. Finally, according to each graph... The system identifies the forwarding behavior attributes of financial data flow direction, transaction relationships, and interaction sequence between nodes and their adjacent upstream and downstream nodes. For each located graph node, a corresponding local data forwarding subgraph is constructed. By first identifying active account nodes with financial data forwarding behavior, then extracting hierarchical attributes to construct a global cascaded topology, and finally generating a single-account subgraph, the system avoids processing invalid data from all static account nodes without transactions. This significantly reduces the overhead of topology construction under massive amounts of financial transaction data. Simultaneously, the cascaded topology graph fully restores the hierarchical link characteristics of financial data flow, and the single-node data forwarding subgraph accurately captures the direct transaction interaction attributes of accounts.

[0082] Step 103: Based on the cascaded topology graph and the data forwarding subgraph, optimize the Euclidean eigenvectors corresponding to each graph node to generate a global hyperbolic embedding matrix;

[0083] Euclidean eigenvectors refer to the initial eigenvectors constructed in Euclidean space based on the attribute features and topological features of graph nodes. They are the basic expression of node features.

[0084] The global hyperbolic embedding matrix is ​​a matrix obtained by mapping all graph nodes to Lorentz hyperbolic space. Each row corresponds to the hyperbolic embedding vector of a node, encoding the node's hierarchical topology and behavioral attributes.

[0085] Please see Figure 3In this embodiment, the global hierarchical association relationship of the cascaded topology graph is used as the global constraint, and the local forwarding interaction relationship of the data forwarding subgraph is used as the local constraint. The initial Euclidean feature vector of each graph node is subjected to Lorentz hyperbolic space exponential mapping projection. Then, based on the mathematical constraints of the Lorentz hyperbolic manifold and the topological association loss function, the projected feature vector is iteratively optimized. This ensures that nodes with data forwarding associations maintain a reasonable geometric distance in hyperbolic space and unrelated nodes are distinguished from each other. In the traffic forwarding scenario, a global hyperbolic embedding matrix of communication nodes is generated. The negative curvature characteristic of hyperbolic space completely solves the problem of the distortion of the representation of hierarchical topology by Euclidean space, realizes the low-dimensional high-fidelity expression of node features, and provides an inductive feature generation basis for dynamically added nodes, improving the generalization and adaptation capability of the embedding matrix.

[0086] In another example, the cascaded topology graph of financial data forwarding is used as the global hierarchical constraint and the data forwarding subgraph is used as the local interaction constraint. The Euclidean eigenvectors of each financial account graph node are subjected to Lorentz hyperbolic space exponential mapping projection. Based on the hyperbolic manifold constraint and the topological association loss function, the projection features are iteratively optimized so that nodes with financial data forwarding associations maintain a reasonable geometric distance in hyperbolic space and unrelated nodes are distinguished from each other. Finally, a global hyperbolic embedding matrix of financial account nodes is generated.

[0087] In one example of the present invention, step 103 may include the following sub-steps:

[0088] Based on the cascaded topology graph and the data forwarding subgraph, create the Euclidean eigenvector corresponding to each graph node in Euclidean space;

[0089] The exponential mapping operator is used to project each Euclidean eigenvector from Euclidean space to hyperbolic space to generate node hyperbolic embedding vectors.

[0090] The initial global hyperbolic embedding matrix is ​​constructed using the hyperbolic embedding vectors of all nodes;

[0091] Using the cascaded topology graph and data forwarding subgraph as constraints, the distance between the hyperbolic embedding vectors of each node in the initial global hyperbolic embedding matrix is ​​adjusted according to the preset hyperbolic distance loss function to generate the global hyperbolic embedding matrix.

[0092] Euclidean space refers to a flat geometric space that satisfies Euclidean geometric axioms. Euclidean eigenvectors are initial feature vectors constructed in Euclidean space based on the topological attributes and forwarding behavior attributes of graph nodes; they are the fundamental representation of node features.

[0093] The exponential mapping operator is the core mathematical operator that realizes the projection of Euclidean space eigenvectors onto hyperbolic space. It can map flat space features to a negative curvature hyperbolic space while preserving the geometric properties of hierarchical topology.

[0094] Hyperbolic space refers to a geometric space with negative curvature. Essentially, the faster a hyperbolic space expands outwards, the more its volume increases exponentially with its radius. It is well-suited for representing hierarchical topologies and can achieve low-distortion representations using low-dimensional embeddings. Hyperbolic graph learning supports three equivalent models: the Poincaré sphere, the Lorentz hyperboloid, and the Klein model.

[0095] The node hyperbolic embedding vector refers to the eigenvector formed by projecting Euclidean eigenvectors onto hyperbolic space using an exponential mapping operator.

[0096] The initial global hyperbolic embedding matrix refers to the matrix formed by arranging all the hyperbolic embedding vectors of all nodes in a regular manner according to the node identifier.

[0097] The hyperbolic distance loss function refers to a loss function constructed based on the geometric rules of hyperbolic space, used to quantify the matching degree between the geometric distance of the hyperbolic embedding vector of a node and the true topological association.

[0098] The global hyperbolic embedding matrix refers to the final matrix after iterative adjustment using the hyperbolic distance loss function, which shows a high degree of matching between the geometric distance of the node hyperbolic embedding vector and the actual topology / forwarding behavior.

[0099] It should be noted that the hyperbolic manifold in the hyperbolic space in this embodiment can be a Lorentz hyperboloid, which is a hyperboloid in (n+1)-dimensional space in which entities can be embedded. Tangent space refers to a local plane approximation near a point. The vector of each entity is ( ), For time components, arrive For spatial components.

[0100] Lorentz inner product formula:

[0101]

[0102] The Lorentz distance is a core metric for measuring the similarity of embeddings of two entities in Lorentz hyperbolic space, and its formula is:

[0103]

[0104] Using the Lorentz model for mapping requires two initializations: ① Entity initialization: First, randomly generate entity vectors in the tangent space, then use exponential mapping to map them to the Lorentz model (hyperbolic space) in one step, obtaining the head entity vector and... Tail entity vector ② Relation initialization: Represent the relation as a parameterized vector. (FHRE) or Lorentz transformation matrix (LorentzKG) et al. The initialization of relations has different specific designs depending on the problem. Then, a scoring function is designed, using the Lorentz distance to measure the similarity between two entity vectors; that is, the smaller the distance, the more relevant the entities. Next, a scoring function is designed to score the quality of unit entity embeddings. Finally, a loss function is designed to summarize the scoring function results for all entities, thereby optimizing the embeddings during subsequent model training. The advantage of this model is that entity embeddings are always within hyperbolic space, avoiding tangent space approximation errors, ultimately achieving low-dimensional, high-precision knowledge graph link prediction. Other hyperbolic KGE models have two fatal problems: First, they require frequent spatial transformations: each training iteration maps entities from hyperbolic space to tangent space (using logarithmic mapping) for computation, and then maps them back to hyperbolic space (using exponential mapping). Each transformation requires calculating complex hyperbolic functions, which is slow and prone to instability. Second, they do not fully utilize hyperbolic space: these models only project entities into hyperbolic space before performing operations, rather than defining operations natively in hyperbolic space, thus failing to leverage the advantages of hyperbolic space.

[0105] In this embodiment, in the scenario of traffic forwarding in a communication network, the cascaded topology graph (representing the global hierarchical transmission law of traffic) and the data forwarding subgraph (representing the local traffic interaction characteristics of a single node) constructed at each time stage are used as the core basis. The network hierarchy attributes, traffic forwarding frequency, number of upstream and downstream links, and other topological and behavioral characteristics of each graph node are integrated in Euclidean space to construct a corresponding Euclidean feature vector for each graph node. The exponential mapping operator adapted to Lorentz hyperbolic space is called to project the Euclidean feature vectors of all graph nodes from Euclidean space to hyperbolic space one by one, and convert them into node hyperbolic embedding vectors that can encode hierarchical topological features. The hyperbolic embedding vectors of all nodes are arranged in order according to the communication node identifier to construct an initial global hyperbolic embedding matrix. With the global association relationship of traffic transmission across levels in the cascaded topology graph and the local association relationship of direct interaction between nodes in the data forwarding subgraph as dual constraints, the preset hyperbolic distance loss function (constructed based on the geodesic distance of Lorentz hyperbolic space) is called to iteratively adjust the geometric distance between the hyperbolic embedding vectors of each node in the initial global hyperbolic embedding matrix. This ensures that nodes with stable traffic forwarding associations maintain a geometric distance that conforms to the hierarchical rules in the hyperbolic space, while nodes without forwarding associations are far apart from each other, and finally, a global hyperbolic embedding matrix is ​​generated. By transforming the initial features from Euclidean space to hyperbolic space projection, the problems of distortion and dimensional redundancy in the hierarchical traffic topology representation of communication networks by Euclidean space are solved. Then, through the constraint optimization of the hyperbolic distance loss function, the geometric distribution of node features within the matrix is ​​highly matched with the topological correlation of real traffic forwarding. This preserves the global hierarchical rules of traffic forwarding and accurately captures the local interaction features of single nodes, providing a high-fidelity, low-dimensional feature carrier for subsequent traffic forwarding target prediction, and greatly improving the accuracy and robustness of prediction.

[0106] In the context of financial data forwarding, the topological and behavioral characteristics of each graph node, such as the depth of financial data flow, the frequency of account transactions, and the number of upstream and downstream transaction associations, are integrated in Euclidean space to construct a corresponding Euclidean feature vector for each graph node. An exponential mapping operator adapted to Lorentz hyperbolic space is invoked to project the Euclidean feature vectors of all graph nodes one by one from Euclidean space to hyperbolic space, converting them into node hyperbolic embedding vectors capable of encoding hierarchical topological features. All node hyperbolic embedding vectors are arranged in order according to financial account identifiers to construct an initial global hyperbolic embedding matrix. Using the global association of financial data flow across levels in the cascaded topology graph and the local association of direct account transactions in the data forwarding subgraph as dual constraints, a preset hyperbolic distance loss function (constructed based on geodesic distance in Lorentz hyperbolic space) is invoked to iteratively adjust the geometric distance between the hyperbolic embedding vectors of each node within the initial global hyperbolic embedding matrix. This ensures that account nodes with stable financial data flow associations maintain a geometric distance consistent with hierarchical rules in hyperbolic space, while nodes without transaction associations are kept far apart, ultimately generating the global hyperbolic embedding matrix. It should be noted that the construction of Euclidean feature vectors can be based on explicit transaction features and contextual structure features. Explicit transaction features capture the instantaneous transaction behavior characteristics of an address, characterizing the tangential coordinates of the address in the hyperbolic map. This includes transfer amount, Gas Price (e.g., high Gas prices for preemptive trades), Gas Limit, transaction timestamp, and contract function address. Contextual structure features characterize the topological location of the address in the global network, characterizing the radial coordinates of the address in the hyperbolic map. This includes the depth of financial data sources (number of hops in financial data flow) and related party attributes (sender's historical activity, centrality, and blacklist risk status).

[0107] Wherein, the hyperbolic embedding vector of the node is:

[0108] ;

[0109] in, Let be the hyperbolic embedding vector of graph node u, and K be a preset manifold curvature parameter. Let be the Euclidean eigenvector of graph node u. For the exponential mapping operator, .

[0110] In the context of traffic forwarding in communication networks, after mapping variables in the topological context features, the first component of the mapping result (determined by cosine) geometrically represents the "depth" of a node in the communication network topology. This is achieved by coupling network hierarchy, traffic access depth, and node centrality into... The model anchors core hub nodes (such as core network elements and data center core switches) at the near end of the time-like axis of the hyperboloid, while a massive number of end nodes (such as terminal devices and IoT nodes) are pushed to the edge of the manifold. After mapping the variables in the inherent attribute characteristics of the nodes, the residual component of the mapping result (determined by sin) defines the functional orientation of the node within the same level. Through the directional differences in traffic behavior patterns and business type characteristics, the model clusters nodes with similar business attributes and traffic characteristics in similar angular sectors of the hyperboloid space, thereby achieving a collaborative representation of node attributes and hierarchical positions.

[0111] Through the aforementioned nonlinear projection, the system constructs all nodes as a hyperbolic embedding matrix E. Under this representation system, the relationships between nodes are no longer measured by linear distance, but by geodesic distance defined by the Minkowski inner product. This process provides a static coordinate reference system with global topology awareness for subsequent diffusion learning, enabling subsequent traffic forwarding pattern learning to be established on a stable hierarchical structure.

[0112] In the scenario of financial data forwarding, the first component of the mapping result (by...) The decision () geometrically characterizes the depth of an address within the financial topology. This is achieved by coupling the depth of financial data sources and activity centrality into... The core hub nodes of the model (such as CEX cold accounts) are anchored at the near end of the hyperboloid's time-like axis, while massive end addresses (such as one-off graph nodes) are pushed to the edge of the manifold. After the variables in the explicit transaction features are mapped, the remaining components of the mapping result (by...) The decision defines the functional orientation of an address within the same level. By leveraging behavioral patterns and the directional differences in Gas features, the model clusters features with similar business attributes in sectors with similar angles in hyperbolic space, thereby achieving a collaborative representation of address attributes and hierarchical positions.

[0113] In another example of the present invention, step 103 may further include the following sub-steps:

[0114] When a new graph node is added, the forwarding behavior attribute associated with the new graph node and the node level attribute it belongs to are extracted, and a new Euclidean feature vector corresponding to the new graph node is created.

[0115] The newly added Euclidean eigenvectors are projected onto the hyperbolic space using the exponential mapping operator to generate hyperbolic embedding vectors for the new nodes.

[0116] Add the node hyperbolic embedding vector to the global hyperbolic embedding matrix to obtain a new global hyperbolic embedding matrix.

[0117] Newly added graph nodes refer to dynamic nodes that are newly connected to the network within a time period and generate data forwarding behavior for the first time, without historical topology and feature records.

[0118] In this embodiment, when a new traffic forwarding graph node appears in the communication network, the traffic forwarding behavior attributes associated with the new node and the network node hierarchy attributes are extracted. Based on this, a corresponding new Euclidean feature vector is constructed. The new Euclidean feature vector is projected onto the hyperbolic space using a predetermined exponential mapping operator to generate a new hyperbolic embedding vector for the new node that is adapted to the hierarchical topology. This embedding vector is then added to the original global hyperbolic embedding matrix to obtain a new global hyperbolic embedding matrix that is compatible with the new node. Dynamic updates can be completed without retraining the entire matrix, effectively solving the cold start problem of new communication nodes and ensuring the real-time performance of the embedding matrix and the consistency of topological representation.

[0119] In another example, when a new financial data forwarding graph node appears in the financial transaction network, the financial data flow behavior attributes associated with the new node and the transaction node hierarchy attributes are extracted. Based on this, a corresponding new Euclidean feature vector is constructed. The new Euclidean feature vector is projected onto the hyperbolic space using a predetermined exponential mapping operator to generate a hyperbolic embedding vector for the new node that is adapted to the transaction link. This embedding vector is then added to the original global hyperbolic embedding matrix to obtain a new global hyperbolic embedding matrix that is compatible with the new account. This allows for rapid adaptation to the new account without re-iteration training, effectively solving the cold start problem of new accounts in financial scenarios that lack historical data, and ensuring the complete representation capability of the embedding matrix for the transaction topology.

[0120] Step 104: Extract the node phase features corresponding to the currently active graph node from the global hyperbolic embedding matrix;

[0121] Currently active graph nodes refer to graph nodes that exhibit traffic forwarding behavior within the current time sequence.

[0122] Node phase features refer to hyperbolic feature vectors extracted from the global hyperbolic embedding matrix, which represent the characteristic state of active graph nodes in the current temporal phase, and integrate node attributes with topological temporal features.

[0123] In this embodiment, based on the identity of active graph nodes in the current time series stage, the hyperbolic embedding vector of the corresponding node is accurately matched and extracted from the pre-generated global hyperbolic embedding matrix. This vector is used as the node's stage feature, thereby directly reusing the pre-trained and optimized high-fidelity features, avoiding the overhead of repeated feature calculations, and ensuring the spatial consistency between the stage features and the global topological features, providing a stable feature foundation for subsequent perturbation updates.

[0124] In another example, based on the financial data forwarding records of the preceding time series corresponding to the current time series stage, the unique identifier of the active graph node that currently generates financial data forwarding behavior is determined. Based on this identifier, the hyperbolic embedding vector of the corresponding node is accurately matched and extracted from the pre-generated and optimized global hyperbolic embedding matrix, which serves as the node stage feature of the active graph node.

[0125] In one example of the present invention, step 104 may include the following sub-steps:

[0126] Extract the initial node features corresponding to the currently active graph nodes from the global hyperbolic embedding matrix;

[0127] The initial node features are averaged and pooled to obtain the node stage features.

[0128] Please see Figure 3 In this embodiment, in the scenario of traffic forwarding in a communication network, based on the traffic forwarding records and cascade topology of the current time series stage, all currently active graph nodes participating in the traffic cascade flow are determined. From the pre-generated and optimized global hyperbolic embedding matrix, the initial node features corresponding to each active graph node are accurately matched and extracted. Then, a hyperbolic space-adapted average pooling operation is performed on the extracted multiple sets of initial node features. The dimensional mean of each dimension of all feature vectors is calculated in accordance with the manifold constraint, and finally, the node stage features that characterize the overall characteristics of traffic forwarding in the current time series stage are obtained. Through this operation, the individual features of multiple scattered single nodes in the current stage are aggregated into a unified low-dimensional stage global feature. This not only fully preserves the core hierarchical rules and behavioral commonalities of traffic forwarding in the current stage, but also effectively filters out redundant noise and individual biases caused by instantaneous traffic fluctuations of single nodes and sudden small packet transmissions. At the same time, it significantly reduces the feature dimension and the computational complexity of subsequent time series modeling.

[0129] In another example, in the scenario of financial data forwarding, based on the transaction flow records and the cascaded topology of financial data in the current time series stage, all currently active graph nodes participating in the cascaded flow of financial data are identified. From the pre-generated and optimized global hyperbolic embedding matrix, the initial node features corresponding to each active graph node are accurately matched and extracted. Then, a hyperbolic space-adapted average pooling operation is performed on the extracted multiple sets of initial node features. The dimensional mean of each dimension of all feature vectors is calculated in accordance with the manifold constraint, and finally, the node stage features that represent the overall characteristics of the financial data flow in the current time series stage are obtained. Through this operation, the individual features of multiple account nodes participating in the financial data flow in the current stage are aggregated into a unified financial data flow stage feature.

[0130] Specifically, the phased characteristics of nodes The generation process can be represented as:

[0131] ;

[0132] in, ( ) represents average pooling. Let be the set of all graph nodes that receive data at time t. This is the feature extraction function.

[0133] In the data flow In the step, based on the generated global hyperbolic embedding matrix Calculate the stage characteristics of graph nodes that receive new data in the current stage. :

[0134] ;

[0135] It's important to note that average pooling is a fundamental technique in graph neural networks for achieving group feature aggregation. It compresses a high-dimensional set of individual features into a low-dimensional group feature vector by averaging the geometric feature dimensions of the target nodes in the graph. For example, an N×d dimensional feature matrix can be transformed into a 1×d matrix through averaging across dimensions. The essence of average pooling is to preserve the common features of the node group while filtering out redundant and unique attributes of individuals.

[0136] Step 105: Update the node stage features by perturbation according to the preset fluctuation perturbation operator to obtain the data prediction feature vector;

[0137] Fluctuation perturbation operators refer to learnable operators built based on probability distributions, used to simulate uncertainties such as sudden fluctuations in network traffic, environmental changes, and trends in financial data flows.

[0138] The data prediction feature vector refers to the feature vector after coupling with perturbation information, which represents the expected evolution characteristics of the data in the next stage.

[0139] Please see Figure 3 In this embodiment, a preset fluctuation disturbance operator is invoked to generate a random disturbance term adapted to the current time series stage. The node stage features and the random disturbance term are weighted and fused together according to a preset weight to update the feature vector, so that the feature vector simultaneously carries the deterministic law of traffic forwarding and the uncertain fluctuation characteristics, thereby obtaining the data prediction feature vector. This simulates the random evolution characteristics of traffic in the real network and improves the robustness of the prediction model to sudden traffic and network fluctuations.

[0140] In another example, a preset fluctuation disturbance operator is invoked to generate a random disturbance term that adapts to the financial transaction environment and financial data flow characteristics of the current time series stage. The extracted phase-specific features of the financial data forwarding nodes are then weighted and fused with this random disturbance term according to a preset fusion rule to complete the feature disturbance update. Finally, a data prediction feature vector corresponding to the financial data forwarding scenario is generated. Through this operation, the generated prediction feature vector simultaneously carries the deterministic evolutionary law determined by the historical behavior of financial data flow, as well as the uncertain characteristics brought about by market environment fluctuations and changes in transaction behavior, perfectly matching the random evolutionary characteristics of data flow in real financial transaction networks.

[0141] In one example of the present invention, step 105 may include the following sub-steps:

[0142] The pre-defined recurrent neural network is invoked to generate the initial data representation corresponding to the current time step based on the stage features of the nodes and the historical hidden states.

[0143] The initial data representation is weighted and fused according to the dynamically adjusted weights and the preset fluctuation disturbance operator to generate the intermediate data representation corresponding to the current time step;

[0144] A nonlinear mapping function is used to nonlinearly fuse the intermediate data representation with the expectation estimation of the fluctuation perturbation operator to obtain the data prediction feature vector.

[0145] Recurrent neural networks (RNNs) are pre-built and trained time-series modeling networks with long-sequence memory effects, used to capture the historical evolution of data forwarding. An example is the Gated Recurrent Unit (GRU).

[0146] The historical hidden state refers to the state vector output by the recurrent neural network at the previous time step, which carries information about the historical evolution of data forwarding. It is the carrier of historical information for the current time step's time series modeling.

[0147] In this embodiment, in the scenario of traffic forwarding in a communication network, a preset recurrent neural network is invoked. The node's stage-specific features at the current time step and the historical hidden state of the previous time step are used as input. The network's gating mechanism captures the historical temporal evolution patterns of traffic forwarding, generating an initial data representation that only carries the deterministic traffic evolution trend at the current time step. Then, based on the stability of traffic forwarding at the current time step, dynamically adjusted weights are determined. According to these weights, the initial data representation is weighted and fused with the random perturbation term output by a preset fluctuation perturbation operator, generating an intermediate data representation that simultaneously carries the deterministic traffic patterns and some fluctuation characteristics. Finally, a nonlinear mapping function is used to integrate the intermediate data representation with the fluctuation... The expectation estimation of the perturbation operator is deeply nonlinearly fused to obtain a data prediction feature vector that simultaneously carries the deterministic evolution law of traffic forwarding and the average trend of uncertainty. Through the temporal memory effect of the recurrent neural network, the evolutionary inertia and link transmission law of traffic across time stages are effectively captured, avoiding the loss of long-sequence temporal information. Then, through adaptive balancing of dynamically adjusted weights, the feature retains the core deterministic trend of traffic forwarding while incorporating the randomness characteristics brought about by network environment fluctuations. Finally, by combining the expectation estimation with the nonlinear mapping function, the predictive ability of the feature for future traffic evolution is further enhanced, fundamentally improving the temporal accuracy and robustness of traffic prediction in sudden scenarios.

[0148] In another example, in a financial data forwarding scenario, a pre-defined recurrent neural network is invoked. The node stage characteristics of the current time series and the historical hidden state of the previous time step are used as input. The network's gating mechanism captures the historical time series evolution of financial data flow, generating an initial data representation that only carries the deterministic financial data flow trend at the current time step. Then, the weights are dynamically adjusted according to the stability of the financial data flow at the current time series. The initial data representation is then weighted and fused with the random disturbance term output by the pre-defined fluctuation disturbance operator according to the weights, generating an intermediate data representation that simultaneously carries the deterministic laws of financial data and some fluctuation characteristics. Finally, a nonlinear mapping function is used to perform deep nonlinear fusion of the intermediate data representation and the expectation estimate of the fluctuation disturbance operator to obtain a data prediction feature vector that simultaneously carries the deterministic evolution law of financial data flow and the uncertain average trend.

[0149] Specifically, a nonlinear mapping function with memory effect is constructed from gated cyclic units, the purpose of which is to calculate the deterministic displacement from stage t to t+1. The initial data representation at the current stage t... for:

[0150]

[0151] Specifically, updating the gate controls the extent to which previous state information is retained in the current state, while resetting the gate allows the model to ignore certain transient information that is not of reference value. For vector concatenation operators, To learn the transition weight matrix for the jump probability between nodes in different types of graphs, This is the hidden state at this time point. Operator Output It successfully captured the deterministic component in the pattern, that is, after excluding random interference, the data flow preference is located based on its historical path and current node attributes, which is the most likely next hop trend.

[0152] It should be noted that the first The fluctuation perturbation operator in the step can be a learnable flow direction offset variable, i.e. .

[0153] Among them, the mean and The average trend and forwarding fluctuation operator representing the data flow direction at that moment are used to characterize the uncertainty of the data forwarding path. The initial data representation is weighted and fused according to dynamically adjusted weights and preset fluctuation disturbance operators to generate the intermediate data representation corresponding to the current time step, which can be expressed as:

[0154] ;

[0155] in, It is a dynamic adjustment of weights, used to adjust the weights of deterministic patterns and uncertain influences on the system.

[0156] A nonlinear mapping function is used to nonlinearly fuse the intermediate data representation with the expectation estimation of the fluctuation perturbation operator to obtain the data prediction feature vector:

[0157]

[0158] in, This is an expected estimate of the disturbance at the next time step. This link combines the nonlinear mapping function learned from previous node flows, nonlinearly fusing the current evolutionary state with the expected random fluctuations, thus completing the prediction for time n+1. This vector describes the geometric position and characteristics the target address should possess in hyperbolic space if the pattern continues to apply.

[0159] Optionally, after performing the step of weighted fusion of the initial data representation according to dynamically adjusted weights and preset fluctuation perturbation operators to generate the intermediate data representation corresponding to the current time step, the following sub-steps are also included:

[0160] The activation function is called to update the hidden state corresponding to the current time step as the new historical hidden state, based on the intermediate data representation and the weight matrix corresponding to the current time step.

[0161] In this embodiment, the intermediate data representation is updated according to the weight matrix corresponding to the current time step, specifically as follows:

[0162] ;

[0163] It is the learnable weight matrix at time step t. It is an activation function.

[0164] Hidden state Continuously accumulate path characteristics (such as the frequency of hierarchical transitions). After... Second recursion, final state It has been internalized as a full lifecycle behavior label that includes complete evolutionary logic and transaction randomness.

[0165] During step 105, the diffusion probability model iterates, with its core operating mechanism consisting of two symmetric Markov chains: forward diffusion and backward diffusion. The forward process distributes the true data... By gradually increasing the noise through a T-step process, the distribution is transformed into a simpler distribution. The diffusion kernel at step t (such as the Gaussian diffusion kernel or the binomial diffusion kernel) is denoted as... The formula for calculating the distribution during Gaussian diffusion is:

[0166] ;

[0167] Where the diffusivity is the diffusivity (Controlling the amount of noise added at each step). The final data distribution is obtained by multiplying the actual data by the diffusion kernel at each step. .

[0168] The back diffusion process is the reverse of learning forward diffusion, from simple noise... Gradually restore the real data The trajectory of the back diffusion is:

[0169] ;

[0170] in, .

[0171] The specific learning process uses multilayer perceptrons and other methods to learn the parameters of the back-diffusion kernel (Gaussian diffusion kernel learns the mean and covariance, binomial diffusion kernel learns the bit-flipping probability), maximizing the lower bound K of the log-likelihood. Furthermore, the diffusion model achieves efficient performance evaluation through techniques such as trajectory probability ratio transformation and diffusion rate reduction. The training objective can also be simplified to the mean squared error of noise prediction and incorporated into time-step embedding and self-attention mechanisms, with a fixed linear growth of the diffusion rate.

[0172] Step 106: Calculate the similarity between the data prediction feature vector and the candidate global hyperbolic embedding matrix, and locate the target prediction graph node with a similarity greater than or equal to the preset similarity threshold.

[0173] The candidate global hyperbolic embedding matrix refers to a hyperbolic embedding matrix that contains all potential traffic receiving nodes, and is a candidate feature library for predicting target nodes.

[0174] Similarity refers to the feature matching degree calculated based on geodesic distance or inner product in Lorentz hyperbolic space.

[0175] In this embodiment, the matching degree between the data prediction feature vector and the embedding vector of each node in the candidate global hyperbolic embedding matrix is ​​calculated respectively. The calculated similarity is compared with a preset similarity threshold. Graph nodes with similarity greater than or equal to the threshold are selected and located as target prediction graph nodes. Thus, accurate target node positioning is achieved based on the geometric matching characteristics of hyperbolic space, ensuring the matching of prediction results with the actual evolution law of traffic.

[0176] In another example, the feature matching degree between the generated financial data forwarding data prediction feature vector and the embedding vector of each candidate node in the candidate global hyperbolic embedding matrix is ​​calculated to obtain the similarity value corresponding to each candidate node. The similarity values ​​of all candidate nodes are compared with the preset similarity threshold one by one, and all graph nodes with similarity greater than or equal to the similarity threshold are selected and located as the target prediction graph nodes for the financial data forwarding scenario.

[0177] In one example of the present invention, the candidate global hyperbolic embedding matrix consists of hyperbolic embedding vectors of multiple candidate graph nodes; step 106 may include the following sub-steps:

[0178] The inner product between the predicted feature vector of the data and the candidate global hyperbolic embedding matrix is ​​calculated as the similarity of each candidate graph node;

[0179] Multiple candidate graph nodes with a similarity greater than or equal to a preset similarity threshold are selected as intermediate graph nodes;

[0180] Select a preset number of intermediate graph nodes as target prediction graph nodes based on their similarity from high to low.

[0181] In the context of traffic forwarding in communication networks, based on the native geometric rules of Lorentz hyperbolic space, the Lorentz inner product between the generated data prediction feature vector and the embedding vector of each candidate graph node in the candidate global hyperbolic embedding matrix is ​​calculated. The result of this inner product is used as the similarity of the corresponding candidate graph node to quantify the matching degree between the candidate node and the predicted traffic evolution features. Then, the similarity of all candidate graph nodes is compared with a preset similarity threshold one by one. Multiple candidate graph nodes with similarity greater than or equal to the similarity threshold are located and selected as intermediate graph nodes. Invalid candidate nodes with low matching degree are eliminated at once to narrow the selection range. Finally, the intermediate graph nodes are sorted in descending order according to the similarity value from high to low, and nodes that match a preset number are selected as the target predicted graph nodes for traffic forwarding. By calculating similarity using the Lorentz inner product, the geometric consistency and computational accuracy of feature matching within hyperbolic space are ensured, while significantly reducing computational overhead compared to hyperbolic geodesic distance calculation. This perfectly meets the real-time requirements of communication network traffic prediction. Furthermore, through a two-step screening mechanism of threshold initial screening and sorting fine selection, redundant nodes with no matching degree are filtered out, while high-probability nodes most likely to become traffic forwarding targets are accurately identified. This effectively improves the accuracy and processing efficiency of traffic forwarding target prediction, providing accurate prediction basis for communication network resource pre-scheduling, early congestion warning, and link load balancing.

[0182] In another example, in the scenario of financial data forwarding, multiple candidate graph nodes with similarity greater than or equal to the similarity threshold are located and filtered as intermediate graph nodes. Irrelevant account nodes with low matching degree are eliminated to compress the target screening range. Finally, the intermediate graph nodes are sorted in descending order according to the similarity value from high to low, and nodes that match a preset number are selected as target prediction graph nodes for financial data forwarding.

[0183] Specifically, the similarity at time step n+1 The formula is expressed as:

[0184]

[0185] in, Let be the candidate global hyperbolic embedding matrix composed of candidate graph nodes. Predict feature vectors for the data.

[0186] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0187] The timing link prediction device based on hyperbolic diffusion network provided in the embodiments of the present invention will be described below. The timing link prediction device based on hyperbolic diffusion network described below can be referred to in correspondence with the timing link prediction method based on hyperbolic diffusion network described above.

[0188] Please see Figure 4 This invention provides a time-series link prediction device based on hyperbolic diffusion networks, comprising:

[0189] The time series partitioning module 401 is used to obtain the time series topology map corresponding to the historical time period and divide it into multiple time series stages;

[0190] The location extraction module 402 is used to locate the graph nodes that exhibit forwarding behavior at each time stage from the time-series topology graph, and extract the cascaded topology graph and data forwarding subgraph corresponding to the graph nodes;

[0191] The global hyperbolic embedding matrix generation module 403 is used to optimize the Euclidean eigenvectors corresponding to each graph node based on the cascaded topology graph and the data forwarding subgraph to generate a global hyperbolic embedding matrix.

[0192] The node stage feature extraction module 404 is used to extract the node stage features corresponding to the currently active graph nodes from the global hyperbolic embedding matrix;

[0193] The data prediction feature vector generation module 405 is used to update the stage features of nodes by perturbation according to the preset fluctuation perturbation operator to obtain the data prediction feature vector.

[0194] The graph node prediction module 406 is used to calculate the similarity between the data prediction feature vector and the candidate global hyperbolic embedding matrix, and to locate the target predicted graph node whose similarity is greater than or equal to a preset similarity threshold.

[0195] Optionally, the positioning extraction module 402 is specifically used for:

[0196] Locate the graph nodes in the time sequence topology graph that exhibit forwarding behavior at each time stage;

[0197] Extract the node hierarchy attributes corresponding to each graph node, and associate the graph nodes to construct a cascaded topology graph;

[0198] Construct a data forwarding subgraph for each graph node based on the forwarding behavior attributes between each graph node and its adjacent graph nodes.

[0199] Optionally, the global hyperbolic embedding matrix generation module 403 is specifically used for:

[0200] Based on the cascaded topology graph and the data forwarding subgraph, create the Euclidean eigenvector corresponding to each graph node in Euclidean space;

[0201] The exponential mapping operator is used to project each Euclidean eigenvector from Euclidean space to hyperbolic space to generate node hyperbolic embedding vectors.

[0202] The initial global hyperbolic embedding matrix is ​​constructed using the hyperbolic embedding vectors of all nodes;

[0203] Using the cascaded topology graph and data forwarding subgraph as constraints, the distance between the hyperbolic embedding vectors of each node in the initial global hyperbolic embedding matrix is ​​adjusted according to the preset hyperbolic distance loss function to generate the global hyperbolic embedding matrix.

[0204] Optionally, the global hyperbolic embedding matrix generation module 403 is further used for:

[0205] When a new graph node is added, the forwarding behavior attribute associated with the new graph node and the node level attribute it belongs to are extracted, and a new Euclidean feature vector corresponding to the new graph node is created.

[0206] The newly added Euclidean eigenvectors are projected onto the hyperbolic space using the exponential mapping operator to generate hyperbolic embedding vectors for the new nodes.

[0207] Add the node hyperbolic embedding vector to the global hyperbolic embedding matrix to obtain a new global hyperbolic embedding matrix.

[0208] Optionally, the node hyperbolic embedding vector is:

[0209] ;

[0210] in, Let be the hyperbolic embedding vector of graph node u, and K be a preset manifold curvature parameter. Let be the Euclidean eigenvector of graph node u. This is an exponential mapping operator.

[0211] Optionally, the node stage feature extraction module 404 is specifically used for:

[0212] Extract the initial node features corresponding to the currently active graph nodes from the global hyperbolic embedding matrix;

[0213] The initial node features are averaged and pooled to obtain the node stage features.

[0214] Optionally, the data prediction feature vector generation module 405 is specifically used for:

[0215] The pre-defined recurrent neural network is invoked to generate the initial data representation corresponding to the current time step based on the stage features of the nodes and the historical hidden states.

[0216] The initial data representation is weighted and fused according to the dynamically adjusted weights and the preset fluctuation disturbance operator to generate the intermediate data representation corresponding to the current time step;

[0217] A nonlinear mapping function is used to nonlinearly fuse the intermediate data representation with the expectation estimation of the fluctuation perturbation operator to obtain the data prediction feature vector.

[0218] Optionally, after performing the step of weighted fusion of the initial data representation according to dynamically adjusted weights and preset fluctuation perturbation operators to generate the intermediate data representation corresponding to the current time step, the data prediction feature vector generation module 405 is further used for:

[0219] The activation function is called to update the hidden state corresponding to the current time step as the new historical hidden state, based on the intermediate data representation and the weight matrix corresponding to the current time step.

[0220] Optionally, the candidate global hyperbolic embedding matrix consists of hyperbolic embedding vectors of multiple candidate graph nodes; the graph node prediction module 406 is specifically used for:

[0221] The inner product between the predicted feature vector of the data and the candidate global hyperbolic embedding matrix is ​​calculated as the similarity of each candidate graph node;

[0222] Multiple candidate graph nodes with a similarity greater than or equal to a preset similarity threshold are selected as intermediate graph nodes;

[0223] Select a preset number of intermediate graph nodes as target prediction graph nodes based on their similarity from high to low.

[0224] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the above-described device and module can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0225] In the several embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0226] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0227] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0228] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A time-series link prediction method based on hyperbolic diffusion networks, characterized in that, include: Obtain the time series topology graph corresponding to the historical time period and divide it into multiple time series stages; From the time-series topology graph, locate the graph nodes that exhibit forwarding behavior in each of the time-series stages, and extract the cascaded topology graph and data forwarding subgraph corresponding to the graph nodes; Based on the cascaded topology graph and the data forwarding subgraph, the Euclidean eigenvectors corresponding to each graph node are optimized to generate a global hyperbolic embedding matrix; Extract the node phase features corresponding to the currently active graph node from the global hyperbolic embedding matrix; The node's phased features are perturbed and updated according to a preset fluctuation perturbation operator to obtain a data prediction feature vector; Calculate the similarity between the data prediction feature vector and the candidate global hyperbolic embedding matrix, and locate the target prediction graph node whose similarity is greater than or equal to a preset similarity threshold.

2. The time-series link prediction method based on hyperbolic diffusion networks according to claim 1, characterized in that, The step of locating the graph nodes that exhibit forwarding behavior at each of the time-series stages from the time-series topology graph, and extracting the cascaded topology graph and data forwarding subgraph corresponding to the graph nodes, includes: Locate the graph nodes in the time sequence topology graph that exhibit forwarding behavior at each of the time sequence stages; Extract the node hierarchy attributes corresponding to each graph node, and associate each graph node to construct a cascaded topology graph; Based on the forwarding behavior attributes between each graph node and its adjacent graph nodes, a data forwarding subgraph corresponding to each graph node is constructed.

3. The time-series link prediction method based on hyperbolic diffusion networks according to claim 1, characterized in that, The step of optimizing the Euclidean eigenvectors corresponding to each graph node based on the cascaded topology graph and the data forwarding subgraph to generate a global hyperbolic embedding matrix includes: Based on the cascaded topology graph and the data forwarding subgraph, create a Euclidean feature vector corresponding to each graph node in Euclidean space; The exponential mapping operator is used to project each of the Euclidean eigenvectors from the Euclidean space to the hyperbolic space to generate node hyperbolic embedding vectors. An initial global hyperbolic embedding matrix is ​​constructed using all the hyperbolic embedding vectors of the nodes; Using the cascaded topology graph and the data forwarding subgraph as constraints, the distance between the hyperbolic embedding vectors of each node in the initial global hyperbolic embedding matrix is ​​adjusted according to a preset hyperbolic distance loss function to generate a global hyperbolic embedding matrix.

4. The time-series link prediction method based on hyperbolic diffusion networks according to claim 3, characterized in that, Also includes: When a new graph node is added, the forwarding behavior attribute associated with the new graph node and the node level attribute it belongs to are extracted, and a new Euclidean feature vector corresponding to the new graph node is created. The newly added Euclidean eigenvectors are projected onto the hyperbolic space using the exponential mapping operator to generate hyperbolic embedding vectors for the new nodes. The node hyperbolic embedding vector is added to the global hyperbolic embedding matrix to obtain a new global hyperbolic embedding matrix.

5. The time-series link prediction method based on hyperbolic diffusion networks according to claim 3, characterized in that, The hyperbolic embedding vector of the node is: ； in, Let be the hyperbolic embedding vector of graph node u, and K be a preset manifold curvature parameter. Let be the Euclidean eigenvector of graph node u. This is an exponential mapping operator.

6. The time-series link prediction method based on hyperbolic diffusion networks according to claim 1, characterized in that, The step of extracting the node phase features corresponding to the currently active graph node from the global hyperbolic embedding matrix includes: Extract the initial node features corresponding to the currently active graph node from the global hyperbolic embedding matrix; The initial node features are averaged and pooled to obtain the node stage features.

7. The time-series link prediction method based on hyperbolic diffusion networks according to claim 1, characterized in that, The step of updating the node's phased features by perturbation according to a preset fluctuation perturbation operator to obtain the data prediction feature vector includes: The preset recurrent neural network is invoked to generate the initial data representation corresponding to the current time step based on the node's stage features and historical hidden state; The initial data representation is weighted and fused according to dynamically adjusted weights and preset fluctuation disturbance operators to generate the intermediate data representation corresponding to the current time step; A nonlinear mapping function is used to nonlinearly fuse the intermediate data representation with the expected estimate of the fluctuation perturbation operator to obtain the data prediction feature vector.

8. The time-series link prediction method based on hyperbolic diffusion networks according to claim 7, characterized in that, After performing the step of weighting and fusing the initial data representation according to dynamically adjusted weights and a preset fluctuation disturbance operator to generate the intermediate data representation corresponding to the current time step, the method further includes: The activation function is invoked to update the hidden state corresponding to the current time step as the new historical hidden state, based on the intermediate data representation and the weight matrix corresponding to the current time step.

9. The time-series link prediction method based on hyperbolic diffusion networks according to claim 1, characterized in that, The candidate global hyperbolic embedding matrix is ​​composed of hyperbolic embedding vectors of multiple candidate graph nodes; the step of calculating the similarity between the data prediction feature vector and the candidate global hyperbolic embedding matrix, and locating the target prediction graph node whose similarity is greater than or equal to a preset similarity threshold, includes: The inner product between the predicted feature vector of the data and the candidate global hyperbolic embedding matrix is ​​calculated as the similarity of each candidate graph node; Multiple candidate graph nodes with a similarity greater than or equal to a preset similarity threshold are identified as intermediate graph nodes; Based on the similarity, a preset number of intermediate graph nodes are selected as target prediction graph nodes, arranged from high to low.

10. A time-series link prediction device based on hyperbolic diffusion networks, characterized in that, include: The time series partitioning module is used to obtain the time series topology map corresponding to the historical time period and divide it into multiple time series stages; The location extraction module is used to locate the graph nodes that exhibit forwarding behavior in each of the time-series stages from the time-series topology graph, and extract the cascaded topology graph and data forwarding subgraph corresponding to the graph nodes; The global hyperbolic embedding matrix generation module is used to optimize the Euclidean eigenvectors corresponding to each graph node based on the cascaded topology graph and the data forwarding subgraph to generate a global hyperbolic embedding matrix. The node stage feature extraction module is used to extract the node stage features corresponding to the currently active graph nodes from the global hyperbolic embedding matrix. The data prediction feature vector generation module is used to update the stage features of the node according to the preset fluctuation perturbation operator to obtain the data prediction feature vector. The graph node prediction module is used to calculate the similarity between the data prediction feature vector and the candidate global hyperbolic embedding matrix, and to locate the target predicted graph node whose similarity is greater than or equal to a preset similarity threshold.

Images