An opportunity network dynamic link prediction method based on a decoupling-free social-spatiotemporal selective state space model

Abstract

The application provides an opportunity network dynamic link prediction method based on a decoupling-free social-spatiotemporal selective state space model, which comprises the following steps: intercepting local physical interaction records in a sliding time window to construct an initial discrete adjacency matrix; introducing a self-loop mechanism for the initial discrete adjacency matrix to generate an augmented local adjacency matrix; calculating a degree matrix based on the augmented local adjacency matrix to construct a symmetric normalized Laplacian matrix; performing real-symmetric eigenvalue decomposition on the Laplacian matrix; quantitatively calculating the social feature missing rate of all network nodes at the current time step; generating an aligned initial representation matrix; outputting spatial enhancement features; constructing a spatiotemporal fusion state transition matrix; introducing a zero-order preservation rule and performing discretization transformation according to a dynamic adaptive step length of input dependence; and predicting the contact edge probability at the future time step through a multilayer perception decoder and an activation function. The prediction accuracy and robustness in an extremely sparse scene are improved at least.

Description

Technical Field

[0001] This invention relates to the fields of data analysis and deep learning processing, and more specifically, to a method for predicting dynamic links in opportunistic networks based on a decoupled social-spatiotemporal selective state-space model. Background Technology

[0002] Opportunistic Networks (OCNs), as an important branch of Mobile Ad Hoc Networks (MAnets), overturn the traditional reliance on continuous end-to-end physical links, instead achieving decentralized communication through an opportunistic "store-carry-forward" mechanism when mobile nodes meet. In this network environment, accurately predicting whether any pair of nodes will make physical contact within a specific future time window is a crucial foundation for optimizing upper-layer mechanisms such as ad hoc routing protocols, distributed cache management, and coordinated network resource scheduling. Unlike conventional wired or cellular networks with relatively stable topologies, nodes in opportunistic networks possess highly free mobility, resulting in link formation with strong randomness, short duration, and highly intermittent contact events. The overall network topology exhibits a stringent environmental constraint characterized by local observability, extreme sparsity, and highly dynamic, interwoven evolution.

[0003] Therefore, how to perform deep implicit association mining on decentralized nodes that lack global topology priors, relying solely on scattered and incomplete local historical contact logs, has become the primary bottleneck in the field of dynamic link prediction. Summary of the Invention

[0004] In view of this, the present invention discloses a dynamic link prediction method for opportunistic networks based on a decoupled social-spatial selective state-space model, which can at least achieve decoupled spatiotemporal coupling evolution, filter long blank periods and output the final link prediction probability, thereby realizing high-precision dynamic link prediction under extreme sparsity and attribute incompleteness conditions.

[0005] Specifically, the present invention is achieved through the following technical solutions:

[0006] This application proposes a dynamic link prediction method for opportunistic networks based on a decoupled social-spatiotemporal selective state-space model. The method is characterized by its applicability to nodes in opportunistic networks lacking global topological priors and exhibiting incomplete heterogeneous social attributes, and includes the following steps:

[0007] The raw contact log stream generated by the mobile node of the receive opportunity network is divided into discrete time steps according to a preset resolution and then processed within a sliding time window. The initial discrete adjacency matrix A(t) is constructed by extracting local physical interaction records.

[0008] A self-loop mechanism is introduced into the initial off-connection matrix A(t), and an identity matrix is ​​added on the diagonal. Generate augmented local adjacency matrix Where N represents the total number of all devices;

[0009] Based on the augmented local adjacency matrix Calculate the degree matrix D(t), and introduce a numerical protection factor into the diagonal elements of the degree matrix D(t). A symmetric normalized Laplace matrix is ​​constructed using a degree matrix with a numerical protection factor. ; for the Laplace matrix Perform real symmetric eigenvalue decomposition, and after non-negativity truncation, select the eigenvectors corresponding to the k smallest eigenvalues ​​as stable low-frequency local spectral features. ;

[0010] By defining an indicator function to scan the original social attributes of the input node by node, the social feature missing rate of all nodes in the network at the current time step is quantitatively calculated. According to the missing rate Adaptive gating activation of learnable missing tokens. The latent semantic distribution behind the missing behavior is automatically learned through gradient backpropagation and implicit completion is performed. Temporary state features are spliced ​​to generate an aligned initial representation matrix X(t).

[0011] The initial representation matrix X(t) is projected through three independent linear mapping layers to generate a query matrix Q(t), a key matrix K(t), and a value matrix V(t). Global homogeneous relationships are then constructed using long-term static social attributes and transformed into a spatial topological mask matrix. The spatial topology mask matrix As a spatial structure bias, it is seamlessly embedded in the scaling dot product self-attention score calculation in an additive form. Utilizing the zero-exponential property of the negative infinity sign in the Softmax operation, it cuts off noise propagation from nodes without social connections, outputting spatially enhanced features. In this context, SA represents that the feature has incorporated socio-semantic priors;

[0012] Capture the diagonal elements of the self-attention weight matrix The nonlinear projection of this matrix is ​​used as a dynamic adjustment factor for the time memory coefficient, and then superimposed on the basic continuous state transition matrix A of the selective state-space model to construct the spatiotemporal fusion state transition matrix. Introduce a zero-order preserve rule and dynamically adaptive step size based on input dependence. Discretization transformation is performed to realize the hidden state of the node. The first-order linear selective recursive update utilizes the intensity of spatial social interaction to adaptively filter lengthy contactless periods and lock key contact events, whereby... It is the social self-attention weight matrix at the t-th sliding time step;

[0013] After undergoing a complete time window recursive evolution, the final coupled hidden state representations of the target prediction node pairs are extracted and concatenated along the channel dimension. The contact connection probabilities for future time steps are then inferred using a multilayer perceptron decoder and a sigmoid activation function. Where u and v represent any two independent individuals in the set of network nodes V. Indicates the current moment. This represents the moment in the future that is immediately following historical observations and yet to be predicted.

[0014] In some embodiments, the Laplace matrix For microscopic construction, the numerical protection factor This is to prevent numerical collapse caused by performing inversion or square root operations on the degree matrix when the degree of the opportunistic network node is zero in the current time slice.

[0015] In some embodiments, the adaptive feature interpolation performs the following steps:

[0016] When the original social attribute vector of the target node is determined by the indicator function When the vector is all zero, the learnable "missing tokens" are forcibly filled at the front end of the feature map;

[0017] When the original social attribute vector of the target node is determined to be non-empty, the social features are mapped to a dense and continuous semantic embedding space through a linear projection layer to eliminate the random multimodal noise introduced by the zero-vector hard-filled directed graph network.

[0018] In some embodiments, at the feature mapping front end of the heterogeneous encoder, for any node i, a feature based on... Adaptive interpolation formula for sociosemantic embedding with dual constraints of local norm:

[0019] ;

[0020] in, The learnable weight matrix for the linear projection layer of social features. For the learnable bias vector of the linear projection layer of social features. Let be the socio-semantic embedding representation of node i at time step t after adaptive interpolation, where t is the current sliding historical time step. , where i is the index of the target node currently being scanned by the interpolation algorithm.

[0021] In some embodiments, the adaptive feature interpolation performs the following steps:

[0022] When the original social attribute vector of the target node is determined by the indicator function When the vector is all zero, the learnable missing tokens are forcibly padded at the front end of the feature map;

[0023] When the original social attribute vector of the target node is not empty, the social features are mapped to a dense and continuous semantic embedding space through a linear projection layer to eliminate the random multimodal noise introduced by the zero-vector hard-filled directed graph network.

[0024] In some embodiments, the spatial topology mask matrix This includes micro-element setting rules, which cause the spatial attention score between nodes without social background to be compressed to zero after exponential normalization.

[0025] In some embodiments, the spatial topology mask matrix Its microscopic elements are mathematically constructed according to a strict step-asymmetric definition:

[0026] ;

[0027] in, It is a matrix The specific element value in the i-th row and j-th column, Represents the dimension of the matrix. Represents the topological mask matrix It is a two-dimensional square matrix consisting of N rows and N columns of real numbers.

[0028] In some embodiments, the construction of the spatiotemporal fusion state transition matrix is ​​used to balance the intensity of dynamic spatial social interaction with the decay rate of inherent temporal historical memory.

[0029] In some embodiments, the dynamic negative sampling strategy performs the following steps:

[0030] For each positive sample edge in the current training batch that has had an actual topological contact A negative sampling set is formed by randomly selecting an equal number of samples from node pairs that have no actual contact records across the entire network. This ensures that the training target set for calculating the binary cross-entropy loss function maintains a 1:1 balance in the positive and negative class distributions.

[0031] In some embodiments, the scale of the sliding time window in the evolutionary sequence is fixed at [specific value]. Each discrete time step has a sampling time resolution between 270 seconds and 3600 seconds, depending on the dynamic change frequency of the decentralized application scenario.

[0032] This application proposes a dynamic link prediction method for opportunistic networks based on a decoupled social-spatiotemporal selective state-space model. First, it extracts local observation subgraphs from the sparse contact logs generated in the opportunistic network by applying a sliding window and performing spectral domain mapping to obtain a stable structural representation. Then, addressing the issue of significant missing social attributes in real-world data, an adaptive interpolation mechanism automatically generates implicit semantic completion of missing attributes. Building upon this, a social-aware space injector is used to inject attention mechanisms based on social homogeneity as a structural bias, bridging physical topological faults. Finally, the extracted spatial social weights are seamlessly embedded into the selective state-space equations, achieving decoupled spatiotemporal coupling evolution, filtering long gaps, and outputting the final link prediction probability. This enables high-precision dynamic link prediction even under conditions of extreme sparsity and incomplete attributes.

[0033] This application introduces a social-aware injection and decoupled state space mechanism into the original architecture of "spatial convolution followed by temporal loop" or parallel extraction, resulting in a novel social-spatiotemporal coupling algorithm to achieve instantaneous entanglement of spatiotemporal characteristics and precise locking of long-range dynamic dependencies. 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 illustrating the dynamic link prediction method for opportunity networks based on a decoupled social-spatiotemporal selective state-space model provided in this application embodiment;

[0036] Figure 2 A schematic diagram of the overall framework of the decoupled social-spatiotemporal selective state-space model provided in this application embodiment;

[0037] Figure 3 A flowchart of local spectral feature extraction provided in an embodiment of this application;

[0038] Figure 4 This is a structural diagram of a socially aware space injector provided in an embodiment of this application. Detailed Implementation

[0039] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of systems and methods consistent with some aspects of this disclosure as detailed in the appended claims.

[0040] The terminology used in this disclosure is for the purpose of describing particular embodiments only and is not intended to be limiting of the disclosure. The singular forms “a,” “the,” and “the” as used in this disclosure and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0041] It should be understood that although the terms first, second, third, etc., may be used in this disclosure to describe various information, such information should not be limited to these terms. These terms are used only to distinguish information of the same type from one another. For example, without departing from the scope of this disclosure, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0042] This application relates to the field of document information extraction, and more specifically, to a method for dynamic link prediction of opportunity networks based on an undecoupled social-spatiotemporal selective state-space model.

[0043] Opportunistic Networks (OCNs), as an important branch of Mobile Ad Hoc Networks (MAnets), overturn the traditional reliance on continuous end-to-end physical links, instead achieving decentralized communication through an opportunistic "store-carry-forward" mechanism when mobile nodes meet. In this network environment, accurately predicting whether any node pair will make physical contact within a specific future time window is a crucial foundation for optimizing upper-layer mechanisms such as ad hoc routing protocols, distributed cache management, and network resource collaborative scheduling. Unlike conventional wired or cellular networks with relatively stable topologies, nodes in opportunistic networks possess highly free mobility, resulting in link formation with strong randomness, short duration, and highly intermittent contact events. The overall network topology exhibits a stringent environmental constraint of local observability, extreme sparsity, and highly dynamic interwoven evolution. Therefore, how to perform deep implicit association mining on decentralized nodes lacking global topology priors, relying solely on scattered and incomplete local historical contact logs, has become the primary bottleneck in the field of dynamic link prediction.

[0044] To manage and capture the spatiotemporal evolution of network topology, existing dynamic link prediction technologies mainly rely on spatial-temporal graph neural networks (GNNs) or continuous-time dynamic graph models (CTDGs). Existing technologies generally follow three research paths when dealing with spatial structure aggregation, time-series recursion, and heterogeneous multimodal semantic fusion, accompanied by a common engineering deployment approach.

[0045] The first category is the Spatial-then-Temporal Decoupled Method. This type of method follows a two-stage sequential processing paradigm of "spatial features first, temporal features later." At each discrete time step snapshot, explicit spatial edge structures within multi-hop neighborhoods are first aggregated using Graph Convolutional Networks (GCNs) or Graph Attention Networks (GATs). Subsequently, the spatial feature matrices generated sequentially at each time step are input into temporal recursive units such as Long Short-Term Memory Networks (LSTMs) or Recurrent Neural Networks (RNNs) to capture the temporal dependencies of topological evolution. The core assumption of this type of method is that the distribution characteristics of the network's spatial structure and the dynamic evolution logic on the time axis can be completely decoupled, independently, and iteratively iterated at the computational level.

[0046] The second category is the Parallel Decoupled Joint Encoding Method. This type of method attempts to overcome the gradient vanishing and timeline lag problems caused by temporal recursion in serial decoupled architectures. It typically constructs completely independent spatial structure feature processing branches (such as those based on lightweight graph convolution or adaptive hypergraphs) and completely independent temporal evolution feature processing branches (such as those based on one-dimensional time series convolution TCN or Transformer self-attention mechanisms) within the model. After both tracks have completed their independent processing, the spatial and temporal features are explicitly aligned at the end of the model through feature matrix concatenation, weighted summation, or the introduction of shallow cross-attention. The core assumption of this type of method is that heterogeneous spatiotemporal multidimensional evolutionary features can be forcibly fused through shallow feature superposition at the end of the model.

[0047] The third category is the Social Attribute Shallow Embedding Method based on static graph reconstruction. This type of method recognizes the limitations of explicit physical topological signal fragmentation and attempts to compensate for structural sparsity by introducing relatively stable long-term social semantic attributes of nodes (such as interests, behavioral patterns, and organizational affiliations). It often utilizes a multilayer perceptron (MLP) to project high-dimensional discrete social attributes into dense vectors, which are then used as static spatial features to initialize the node feature matrix of the graph neural network. Alternatively, it calculates static social homogeneity similarity and uses it as a globally fixed weighted edge matrix, directly performing shallow addition reconstruction with the dynamic physical adjacency matrix. The core assumption of this type of method is that heterogeneous social semantics are absolutely complete and without missing elements in the spatial dimension, and can be directly used as strong priors in graph topology modeling.

[0048] However, under the hard constraints of local observability and compromised data integrity in opportunistic networks, the three forms mentioned above each suffer from the following irreconcilable technical defects and coordination failures.

[0049] First, the serial decoupled spatiotemporal network method disrupts the instantaneous joint distribution of spatiotemporal variables. Because opportunistic network topologies exhibit typical step-like fluctuations, the frequency of spatial structure changes is itself a time-dependent function, which is artificially fragmented by the serial architecture. When dealing with the typical "long-term no-contact gap—short-term sudden encounter" pattern in opportunistic networks, the spatial module is completely unaware of the memory decay caused by the time span when calculating convolutions. This leads to sparse but crucial sudden encounter signals being easily diluted and forgotten by cyclic temporal units when entering subsequent lengthy time gaps, resulting in severe long-term dependency breaks.

[0050] Secondly, the parallel decoupled spatiotemporal multitrack method lacks instantaneous coordination at the underlying physical computation level. This path allows the temporal branch and the spatial topology branch to proceed in independent semantic flows, making it impossible for spatial features to adaptively adjust according to the real-time step size of temporal evolution, and for temporal features to be unable to perceive real-time spectral domain fluctuations in the local topology. This decoupling artificially removes the dynamic spatiotemporal link in the feature generation stage, and the shallow feature splicing or shallow weighting of the terminal modules cannot fundamentally characterize the highly intertwined, decoupled spatiotemporal deep evolution pattern, resulting in a weak dynamic link state representation in the final generated model.

[0051] Third, shallow embedding methods for social attributes based on static graph reconstruction suffer from large-scale prediction blind spots in scenarios with missing attributes. These methods blindly pursue the unconstrained compensation effect of static social semantics, lacking attention to robust modeling of attribute-missing conditions. When encountering large-scale, high-density random attribute missingness in real-world scenarios, these methods lack adaptive implicit completion channels. This not only fails to effectively restore the potential homogeneous similarity associations between nodes, but also transmits a large number of "dead nodes" and high-dimensional zero-filling white noise to the graph neural network, directly causing feature alignment failure and creating prediction blind spots for potential contact points that are "physically topologically broken but socially semantically similar."

[0052] Fourth, the lack of a coordinated combination mechanism can lead to severe system-level coordination failures when there is a lack of information feedback channels. This is the core problem that this invention directly addresses and solves. Specifically, the spatial feature extraction mechanism (as shown in the self-attention module) and the temporal tracking mechanism (such as the loop or streaming model) execute independently, resulting in severe "information asymmetry." When calculating spatial attention, the spatial injection mechanism is completely unaware of the temporal stability of the topology in the spectral domain, leading to the continuous propagation of interference noise caused by large areas of zero-filling, squeezing out the representation space of truly homogeneous and active nodes. When performing state decay, the temporal tracking mechanism is completely unaware of the potential interaction strength of the current node at the social space level, causing the system to blindly perform memory decay even when strong implicit social interactions occur (at which point explicit physical connections break), resulting in the unnecessary discarding of sparse encounter signals that could guide link prediction. The direct consequences of the above coordination failures are the loss of key evolutionary features, the breakage of long-range dependencies, and the extreme vulnerability of model performance to attribute deficiencies, causing the learning trajectory to reach the fitting bottleneck prematurely.

[0053] Therefore, under the constraints of decentralized networks with local observability and incomplete data, existing technologies still lack a decoupled social-spatial dynamic link prediction and coordination scheme that can deeply constrain the spatial structural features and temporal series features through explicit information channels (such as directly embedding spatial social attention weights into the continuous time state transition matrix) to achieve instantaneous entanglement, while preserving the independent representation mechanism of spatial structural features and temporal series features.

[0054] Based on this, this application proposes a dynamic link prediction method for opportunity networks based on an undecoupled social-spatial selective state-space model.

[0055] Specifically, the method is applied to opportunistic network nodes that lack global topological priors and have incomplete heterogeneous social attributes.

[0056] Please see Figure 1 This application discloses a flowchart illustrating a dynamic link prediction method for opportunity networks based on a decoupled social-spatiotemporal selective state-space model. For example... Figure 1The method for dynamic link prediction in opportunity networks based on a decoupled social-spatiotemporal selective state-space model, as shown, includes:

[0057] S101. Receive the raw contact log stream generated by the mobile node of the opportunistic network, divide it into discrete time steps according to a preset resolution, and within a sliding time window... The initial discrete adjacency matrix A(t) is constructed by extracting local physical interaction records.

[0058] This ensures the self-transfer path of features for isolated nodes in extremely sparse scenarios.

[0059] The scale of the sliding time window within the evolutionary sequence is fixed at [specific value]. Each discrete time step has a sampling time resolution between 270 seconds and 3600 seconds, depending on the dynamic change frequency of the decentralized application scenario.

[0060] S102. Introduce a self-loop mechanism to the initial off-link matrix A(t) and add an identity matrix on the diagonal. Generate augmented local adjacency matrix .

[0061] Where N represents the total number of all devices currently involved in the search and rescue efforts in the disaster area, including search and rescue drones, search and rescue vehicles, handheld terminals for rescue personnel, and mobile devices for survivors. It is The square matrix has all elements on the main diagonal set to 1, and all other elements set to 0.

[0062] S103, Based on the augmented local adjacency matrix Calculate the degree matrix D(t), and introduce a numerical protection factor into the diagonal elements of the degree matrix D(t). A symmetric normalized Laplace matrix is ​​constructed using a degree matrix with a numerical protection factor. ; for the Laplace matrix Perform real symmetric eigenvalue decomposition, and after non-negativity truncation, select the eigenvectors corresponding to the k smallest eigenvalues ​​as stable low-frequency local spectral features. .

[0063] In related technologies, multimodal or heterogeneous feature fusion techniques commonly employ "static hard padding" (such as zero padding or mean imputation) or "end-to-end discard / masking mechanisms" (such as masked language modeling) when faced with missing attributes. Zero padding introduces significant "semantic white noise" into graph networks. Because zero vectors erase the feature heterogeneity of nodes during linear mapping and spatial propagation in graph neural networks (GNNs), subsequent spatial self-attention mechanisms experience severe baseline shifts when calculating dot product similarity, causing heterogeneous feature alignment to completely fail.

[0064] This application proposes a data-driven, dynamically decoupled feature-missing state-sensing imputation channel (SAIM). Instead of performing isolated node-level imputation, we consider the overall network-wide macroscopic missing rate. Under a global awareness, the activation paths for the "data-present" and "data-free" modalities are completely separated. When a node attribute is missing, gating forces it to switch to a dedicated path using a learnable missing token. It acts as a latent semantic fitting channel, completely cutting off the path for zero noise to enter the normal feature manifold.

[0065] Specifically, the Laplace matrix For microscopic construction, the numerical protection factor This is to prevent numerical collapse caused by performing inversion or square root operations on the degree matrix when the degree of the opportunistic network node is zero in the current time slice.

[0066] S104. By defining an indicator function, scan the original social attributes of the input node by node, and quantify the social feature missing rate of all nodes in the network at the current time step. According to the missing rate Adaptive gating activation of learnable missing tokens. The underlying semantic distribution behind the missing behavior is automatically learned through gradient backpropagation and implicit completion is performed. Temporary state features are concatenated to generate an aligned initial representation matrix X(t).

[0067] The adaptive feature interpolation process involves the following steps:

[0068] When the original social attribute vector of the target node is determined by the indicator function When the vector is all zero, the learnable "missing tokens" are forcibly filled at the front end of the feature map;

[0069] When the original social attribute vector of the target node is determined to be non-empty, the social features are mapped to a dense and continuous semantic embedding space through a linear projection layer to eliminate the random multimodal noise introduced by the zero-vector hard-filled directed graph network.

[0070] Specifically, at the feature mapping front end of the heterogeneous encoder, for any node i, a feature based on... Adaptive interpolation formula for sociosemantic embedding with dual constraints of local norm:

[0071] ;

[0072] in, The learnable weight matrix for the linear projection layer of social features. For the learnable bias vector of the linear projection layer of social features. Let be the socio-semantic embedding representation of node i at time step t after adaptive interpolation, where t is the current sliding historical time step. , where i is the index of the target node currently being scanned by the interpolation algorithm.

[0073] The adaptive feature interpolation process involves the following steps:

[0074] When the original social attribute vector of the target node is determined by the indicator function When the vector is all zero, the learnable missing tokens are forcibly padded at the front end of the feature map;

[0075] When the original social attribute vector of the target node is not empty, the social features are mapped to a dense and continuous semantic embedding space through a linear projection layer to eliminate the random multimodal noise introduced by the zero-vector hard-filled directed graph network.

[0076] At time step t, this application first constructs an indicator function. The original social attribute moments are scanned node by node. To rigorously and accurately determine whether node i belongs to a state where all attributes are missing due to privacy concerns or malfunctions, this invention utilizes... Norm replacement of conventional logical NOT gates. The overall social feature missing rate across the internet. The calculation formula is:

[0077]

[0078] This formula, in the runtime environment, directly transforms discrete missing behavior into a range. Continuous global environmental perception metrics between them.

[0079] It should be noted that the token learning rate in related technologies is globally fixed. However, this invention uses MissingToken. A dynamic gradient scaling factor was introduced. (in (These are preset hyperparameter coefficients). When the overall network attribute missing rate... When the threshold is exceeded (e.g., 0.15), the gating is forcibly amplified. This allows the gradient update step size of the Missing Token to be adaptively adjusted during the current batch training.

[0080] This forces the model to accelerate performance even in extremely harsh data environments. By fitting the unobserved, latent social background manifold of the entire network, we elevate it from a simple "data completion problem" to a "data-driven implicit social manifold reconstruction," completely eliminating the noise interference from zero-filling.

[0081] In the feature mapping front end of the heterogeneous encoder, this invention designs a feature mapping method based on arbitrary node i. Adaptive interpolation formula for sociosemantic embedding with dual constraints of local norm:

[0082]

[0083] When the condition is met, the semantic embedding of the missing node is directly replaced with... It completely skips the part containing the learnable weight matrix. and bias vector The linear projection layer. This means that missing nodes no longer participate in the linear spatial mapping of normal nodes, thus cutting off the zero-noise pair. Gradient pollution during parameter updates. Here... It is not a hard-coded static placeholder constant, but a continuous, dense representation vector that is completely adaptively optimized with the system's backpropagation. Under the constraints of the loss function, the gradient flow is automatically driven by backpropagation. To approximate and fit the underlying homogeneous social distribution behind these missing nodes.

[0084] S105. Project the initial representation matrix X(t) through three independent linear mapping layers to generate a query matrix Q(t), a key matrix K(t), and a value matrix V(t). Then, construct a global homogeneous relationship using long-term static social attributes and transform it into a spatial topological mask matrix. The spatial topology mask matrix As a spatial structure bias, it is seamlessly embedded in the scaling dot product self-attention score calculation in an additive form. Utilizing the zero-exponential property of the negative infinity sign in the Softmax operation, it cuts off noise propagation from nodes without social connections, outputting spatially enhanced features. .

[0085] The subscript SA at the bottom indicates that the feature has incorporated socio-semantic priors.

[0086] Wherein, the spatial topology mask matrix This includes micro-element setting rules, which cause the spatial attention score between nodes without social background to be compressed to zero after exponential normalization.

[0087] Specifically, spatial topological mask matrix Its microscopic elements are mathematically constructed according to a strict step-asymmetric definition:

[0088] ;

[0089] in, It is a matrix The specific element value in the i-th row and j-th column, Represents the dimension of the matrix. Represents the topological mask matrix It is a two-dimensional square matrix consisting of N rows and N columns of real numbers.

[0090] It should be noted that the traditional Graph Attention Network (GAT) or Transformer self-attention mechanism calculates attention scores entirely based on data (determined purely by the dot product of Q and K), or uses an explicit topological structure as a fixed mask (1 for edges, 0 for no edges).

[0091] In related technologies, opportunistic networks exhibit extremely fragmented and highly intermittent explicit connections. If relying solely on explicit topological masks, the model will completely lose feature propagation paths (creating prediction blind spots) when connections are broken; if no masks are added and the network is fully connected, noise from large areas of zero-filling will propagate throughout the entire network.

[0092] This application creatively designs an asymmetric spatial topological mask matrix with a negative infinity sign penalty. As a "structural bias," we do not blindly trust current explicit edges, but instead utilize a long-term, relatively stable set of static social attributes. The intersection of these points serves as an implicit bridge.

[0093] The spatial topology mask matrix of the present invention Its microscopic elements are mathematically constructed according to a strict step-asymmetric definition:

[0094]

[0095] Traditional attention mechanisms use topological information as feature input. This invention, however, directly embeds it as an additive bias at the physical layer:

[0096]

[0097] In the subsequent Softmax exponent calculation, This means that as long as two nodes have no overlap in long-term social attributes, regardless of their current feature vectors... No matter how high the calculated similarity is, this invention forcibly compresses its attention score to an absolute zero.

[0098] In this way, the flow direction of higher-order features can be firmly locked onto the implicit bridge induced by "social homogeneity". It successfully cuts off the spread of high-dimensional topological noise caused by attribute loss throughout the network, and achieves the technical goal of accurately capturing potential contact patterns across the fault lines even when the explicit topological connections are completely broken.

[0099] S106. Intercept the diagonal elements of the self-attention weight matrix. The nonlinear projection of this matrix is ​​used as a dynamic adjustment factor for the time memory coefficient, and then superimposed on the basic continuous state transition matrix A of the selective state-space model to construct the spatiotemporal fusion state transition matrix. Introduce a zero-order preserve rule and dynamically adaptive step size based on input dependence. Discretization transformation is performed to realize the hidden state of the node. The first-order linear selective recursive update utilizes the intensity of spatial social interaction to adaptively filter out lengthy contactless gaps and lock key contact events.

[0100] in, It is the social self-attention weight matrix at the t-th sliding time step, that is, the implicit spatiotemporal interaction intensity correlation network among all rescue / search and rescue terminals in the disaster area at the current moment, driven by the social homogeneity background.

[0101] Here, the construction of the spatiotemporal fusion state transition matrix is ​​used to balance the intensity of dynamic spatial social interaction with the decay rate of inherent temporal historical memory.

[0102] It should be noted that in the traditional field of dynamic graph or spatiotemporal sequence modeling, existing technologies for the joint capture of temporal and spatial features mainly follow the following two paths, but neither can adapt to the harsh environment of extremely sparse opportunistic network topologies, highly intermittent contacts, and long blank periods:

[0103] First, the essential difference from the "spatiotemporal decoupling" architecture: The core assumption of traditional serial decoupling (such as EvolveGCN, GC-LSTM, TGCRN) or parallel decoupling (such as STD-MAE, STH-SepNet) architectures is that the temporal evolution dynamics and spatial topology are independent of each other. In terms of computation, this is manifested as "first extracting spatial features through graph convolution, and then serially inputting the feature matrix into a recurrent network (RNN / LSTM)" or "dual-track parallel encoding followed by shallow concatenation at the end".

[0104] The disadvantages of spatiotemporal decoupling: This fragmentation mechanism disrupts the joint distribution of spatiotemporal variables. Because the dynamic spatiotemporal link is artificially severed, the time-cycle unit experiences a disconnection window spanning several weeks across the opportunistic network (…). When the historical memory decay rate is still forced to be maintained globally, the sparse but key encounter features are rapidly diluted and forgotten in the long time step recursion (resulting in severe long-range dependency breaks and memory decay).

[0105] Secondly, the fundamental difference from "weakly coupled / event-driven" architectures: Some weakly coupled dynamic graph models that have emerged in recent years (such as TGN, TGAT, TCL, GraphMixer, and DyGFormer in the continuous-time dynamic graph field) attempt to establish streaming state tracking within a unified framework. They typically explicitly design time decay as a function of "time difference". "The relevant exponential or self-attention time encoding is used to passively trigger the streaming update of the node memory module when a network event arrives."

[0106] Disadvantages of weak coupling: These weakly coupled models are inherently highly dependent on continuous, dense streams of interactive events. In the typical opportunistic network scenario of "long-term disconnections and sporadic, sudden encounters," the streaming tracing flow is prone to interruption or stagnation due to the long-term lack of explicit edge events to trigger state updates. More critically, its temporal memory decay rate is only related to the length of time span, and computationally it remains non-linearly disconnected from spatial dynamics, making it impossible to "intelligently" control the retention or release of memories based on the real-time strength of spatial social connections.

[0107] Compared to existing technologies, this invention proposes a completely decoupled spatiotemporal instantaneous entanglement evolution mechanism. The system no longer passively relies on event-triggered updates, nor does it divide spatiotemporal processes into two stages. This invention creatively directly intercepts the diagonal elements of the spatial social self-attention weight matrix (the social spatial interaction activity of the nodes themselves) and embeds them as adjustment factors into the continuous system time transition matrix of the state-space model (Mamba). In this context, at the very bottom of physical calculations, the "memory retention coefficient of the time dimension" becomes an instantaneous function of "spatial social activity".

[0108] In some optional embodiments, the social self-attention weight matrix containing topological constraints in the space output by the (SAS Injector) is first captured in real time. To accurately quantify the implicit encounter activity intensity of a single node at the current moment driven by social homogeneity, this invention extracts the diagonal elements of the attention matrix. The magnitude of this diagonal value directly reflects the intensity of implicit spatial co-occurrence of the node with a group of nodes that have highly similar social backgrounds within the current sliding window.

[0109] This invention modifies the underlying dynamic differential equations of a continuous linear system carrying time-series recursion, transforming the system's original inherent continuous state transition matrix. Upgraded to a fusion state transition matrix with deeply coupled spatial social priors :

[0110]

[0111] in, This operation is used to extract the self-attention level of nodes on the diagonal of the attention matrix. These are learnable gating parameters used to balance dynamic spatial interactions with inherent time decay. This formula has profound physical implications: if a node experiences strong social spatial interaction (high attention weight) at the current time step, the system will dynamically suppress the matrix. The decay effect allows the memory of this key interaction to be preserved for a longer period of time on the timeline.

[0112] To apply continuous systems to discontinuous discrete contact map snapshot sequences, an adaptive time step is introduced into the system. This makes it dependent on the current input features (spatial enhancement sequence smoothed by one-dimensional convolution), Subsequently, using the zero-order preservation rule, the continuous matrix that deeply integrates spatial priors was... By performing a rigorous discretization transformation, the discrete system matrix that can be directly executed in discrete time steps on a computer is obtained. :

[0113]

[0114]

[0115] By feeding the state transition matrix containing the spatial self-interest factor into In matrix exponential operations, deep, irreversible, instantaneous entanglement of spacetime properties is achieved at the mathematical and physical level.

[0116] In this way, this application overcomes the problems of long-range dependency breakage caused by the fragmented joint distribution in the traditional "spatiotemporal decoupling architecture", as well as the bottlenecks of state stagnation and gradient vanishing that the traditional "weakly coupled / event-driven model" is prone to when there is a lack of dense event flow triggering.

[0117] S107. After undergoing a complete time window recursive evolution, the final coupled hidden state representation of the target prediction node pair is extracted and concatenated along the channel dimension. The contact connection probability of the output future time step is inferred through a multilayer perceptron decoder and a sigmoid activation function. .

[0118] Where u and v represent any two independent individuals in the set of network nodes V. Indicates the current moment. This represents the future moment that is immediately following historical observations and yet to be predicted.

[0119] The dynamic negative sampling strategy performs the following steps:

[0120] For each positive sample edge in the current training batch that has had an actual topological contact A negative sampling set is formed by randomly selecting an equal number of samples from node pairs that have no actual contact records across the entire network. This ensures that the training target set for calculating the binary cross-entropy loss function maintains a 1:1 balance in the positive and negative class distributions.

[0121] Furthermore, it should be noted that this application can be applied to mobile ad hoc networks, vehicle-to-everything (V2X) roadside unit collaborative systems, post-disaster emergency communication networks, or digital smart pastoral area network monitoring environments to provide high-precision node encounter probability decision support for underlying intelligent routing and forwarding, dynamic cache clearing, and intelligent copy deletion mechanisms.

[0122] Please see Figures 1-4 In some embodiments, this application may first dynamically contact network data preprocessing and sliding time window local observation subgraph construction.

[0123] Thus, in the preliminary data preparation stage of the entire link prediction model, the core task is to extract local interaction topology under decentralized constraints, avoiding the introduction of unrealistic global topological priors. First, the raw contact logs generated by mobile nodes in the opportunistic network within a continuous time stream are received, and discrete time steps are divided according to a preset time resolution to construct a sliding time window. Within this window, only the edges between node pairs that have actually had physical contact are retained, generating a basic local observation adjacency matrix. To address the problem of highly sparse nodes and the high likelihood of isolated nodes with zero degree in opportunistic networks, a self-loop mechanism is forcibly introduced into the basic adjacency matrix. Specifically, an identity matrix is ​​added to the diagonal of the adjacency matrix to construct an augmented local adjacency matrix. This operation ensures that the features of isolated nodes can self-circulate in subsequent operations, avoiding singularity and computational collapse of the feature matrix in subsequent calculations, and providing an absolutely safe topological carrier for subsequent spectral domain decomposition.

[0124] Secondly, local spectral features can be extracted based on the symmetric normalized Laplacian matrix with a numerical protection factor.

[0125] Here, the extremely sparse and fragmented discrete topological graph obtained in the previous step is mapped into smooth, dense, and low-dimensional numerical structural anchors. First, the node degree matrix is ​​calculated based on the augmented local adjacency matrix. To prevent numerical instability such as division by zero or gradient explosion when inverting or taking the square root of the degree matrix in extremely sparse scenarios, a very small numerical protection factor is uniformly added to the diagonal elements of the degree matrix. Subsequently, the augmented adjacency matrix is ​​symmetrically normalized using this degree matrix with the protection factor to construct a secure Laplace matrix. Next, eigenvalue decomposition is performed on this real symmetric Laplace matrix. Considering that the truncation of floating-point precision on the computing device may produce extremely small negative eigenvalues, a non-negative truncation operation is performed on all decomposed eigenvalues, that is, all eigenvalues ​​less than zero are forcibly set to zero to ensure the physical rationality of the spectral energy. Finally, the eigenvector corresponding to the smallest preset number of eigenvalues ​​is selected as the local spectral feature of the current time step, and the low-frequency signal is used to characterize the gently changing and more stable basic structural pattern in the graph.

[0126] Secondly, it can perform sparse-aware adaptive attribute imputation and preliminary alignment of heterogeneous features for scenarios with high missing rates.

[0127] Here, the core objective of this step is to address the widespread loss of social attributes in real-world opportunity networks due to privacy restrictions, preventing traditional zero-filling strategies from introducing severe semantic noise into the model. First, an indicator function is defined to scan the original social attribute matrix node-by-node, quantifying the social feature loss rate of all nodes in the network at the current time step. Then, a sparsity-aware gating is set at the front end of the feature mapping network. When a node's original attribute vector is empty, a set of learnable "missing tokens" with a preset dimension is adaptively activated based on the calculated global loss rate. During gradient optimization in the model's backpropagation, these missing tokens automatically learn and fit the underlying homogeneity distribution behind nodes with missing attributes, completing data-driven implicit semantic completion. Finally, the adaptively imputed social semantic embedding is cross-fused with the node's current local observation temporary state along the feature dimension to generate a unified high-dimensional initial node representation, completing the initial spatial alignment of heterogeneous modal information.

[0128] Secondly, a social perception spatial structure bias injection mechanism based on social homogeneity masking is implemented.

[0129] This step is used to overcome explicit topological discontinuities, forcing features to flow along implicit social connections. First, the high-dimensional node representations generated in the previous step are projected onto a query matrix, a key matrix, and a value matrix through three independent linear mapping layers. Simultaneously, relatively stable discrete social attributes from the dataset are extracted to construct a global implicit social network, which is then transformed into a topological mask matrix. The rule is that if nodes share common social attributes, the mask value is zero; otherwise, it is negative infinity. When calculating the attention weights of the scaled dot product of the query and key matrices, this topological mask matrix is ​​directly injected as a structural bias into the attention score in an additive manner. Subsequently, normalization is performed using the Softmax function. Since the negative infinity mask bits are compressed to zero, the model forcibly cuts off noise propagation between nodes with no social connections, ensuring that the final attention weights only reflect the strength of connections based on social homogeneity. Finally, this fused weight is multiplied by the value matrix, and the fused weight is simultaneously separated and passed to the next evolutionary module, outputting spatially enhanced features injected with social priors.

[0130] Subsequently, the temporal transition matrix for spatial attention modulation is constructed and the selective state space is discretized and evolved.

[0131] This step is the core control link for achieving decoupled co-evolution of spatiotemporal characteristics and filtering long periods of inactivity. First, the diagonal element of the spatial social attention weight matrix generated in the previous step during social perception space injection is captured. This element accurately reflects the self-attention intensity of a node's social interaction at the current moment. This attention weight is then directly superimposed onto the basic temporal state transition matrix of the selective state space model after nonlinear projection and adjustment of learnable gating parameters, constructing a highly coupled "spatiotemporal fusion state transition matrix." This mechanism achieves instantaneous entanglement of spatiotemporal variables at the physical computation level: when nodes engage in strong social interaction, the retention coefficient of the transition matrix is ​​dynamically amplified, adaptively suppressing memory decay. Subsequently, a zero-order preservation rule is introduced, and a rigorous discretization operation is performed based on the spatiotemporal fusion state transition matrix, recursively updating the hidden state of nodes at discrete time steps. This micro-design enables the system to adaptively and rapidly traverse lengthy periods of no contact, firmly locking sparse but crucial encounter events into continuous hidden state memory.

[0132] Finally, cross-modal coupled feature decoding and binary cross-entropy loss optimization based on a dynamic negative sampling strategy are performed.

[0133] This step is the final feature output and gradient backpropagation control stage of the entire method, ensuring the convergence effectiveness of the prediction network under extreme class imbalance. First, after the recursive evolution throughout the entire sliding time window, the final high-order spatiotemporal coupling representation of the node pair to be predicted at the target time step is extracted. The representation vectors of these two nodes are precisely concatenated along the channel dimension to generate a joint interaction feature. This interaction feature is input into the multilayer perceptron decoder for channel dimensionality reduction, and the final output logistic value is strictly constrained to a probability range of 0 to 1 using the Sigmoid activation function, serving as the predicted probability of the node pair forming an edge. During model training, considering the extreme sparsity characteristic where the number of actually contacting positive samples is far less than the number of uncontacting negative samples, a dynamic negative sampling strategy is designed. That is, for each real edge in the current batch, a strictly equal number of negative samples are randomly selected from all uncontacting node pairs in the network to form the training set. Finally, based on this strictly balanced sample set, the binary cross-entropy loss is calculated, and all feature mapping weights and state space gating parameters are updated in reverse using a gradient descent optimizer until the model converges.

[0134] In one embodiment, this application provides a specific implementation application using a post-disaster emergency communication scenario as an example. At the disaster site, due to the complete shutdown of ground base stations, a heterogeneous opportunistic network system is formed, characterized by localized observation, extreme sparseness, and highly intermittent contact time, comprised of rescue drones, search and rescue vehicles, handheld terminals of rescue personnel, and mobile devices of disaster survivors. This embodiment elaborates on the specific application scenario of "self-organized adaptive communication for emergency rescue in environments with severe infrastructure damage after earthquakes and floods."

[0135] Specifically, during the movement of search and rescue drones or vehicles in the disaster area, due to communication distance limitations, it is impossible to obtain the global network topology of the entire disaster area (lacking global prior knowledge). Each terminal node only receives and records Bluetooth and WiFi short-range scan contact log streams with other devices in real time locally. A local historical encounter map of the past hour is extracted with a step size of 10 minutes. Due to the extremely dispersed distribution of people in the disaster area, many terminals are in an isolated state with zero connections for most of the time. The algorithm forcibly superimposes a self-loop mechanism in the local adjacency matrix and introduces a numerical protection factor. A symmetric normalized Laplace matrix is ​​constructed, and eigenvalue decomposition and non-negative truncation in the spectral domain are performed on it. This step maps extremely fragmented and scattered short-term instantaneous contact events into numerically stable low-frequency spatial structure anchor points, avoiding numerical collapse of inversion values ​​in vehicle-mounted and drone computing chips caused by isolated nodes, and providing a secure underlying carrier for subsequent search of implicit rescue communication paths.

[0136] Subsequently, in rescue coordination, nodes typically carry relatively stable discrete social attributes (such as: affiliated rescue team, search and rescue function, type of supplies carried, action preferences, and movement trajectory characteristics). However, in the harsh post-disaster environment, due to insufficient battery power, malfunctions, or privacy restrictions on survivor devices, the social attributes of all nodes in the network are often largely missing. If traditional zero-filling is used to fill in the missing attributes, the system will blindly judge missing devices as "dead nodes," thus omitting them when selecting the next hop node.

[0137] This invention quantifies the macroscopic missing rate of all network attributes in the current batch by defining an indicator function. When the missing rate triggers a gating threshold, a learnable "Missing Token" is automatically activated. During rescue missions, the missing token is adaptively fitted and reconstructed using a backpropagation algorithm to fit the underlying social affiliation and behavioral patterns behind the missing attributes of the devices. This completely isolates the semantic noise caused by zero-filling and achieves robust alignment of heterogeneous features in the spatial dimension.

[0138] Secondly, the large-scale disruption of the physical topology in the disaster area (such as two search and rescue teams being blocked by mountains and unable to make contact in a short period of time) has resulted in a serious "prediction blind spot" in traditional networks that rely solely on visible connections for message transmission.

[0139] Here, the present invention extracts the relatively stable long-term social attributes of each rescue node, constructs a global homogeneous relationship, and transforms it into a spatial topological mask matrix. The micro-rules are strictly defined as follows: if two rescue terminals share common social attributes (such as belonging to the same rescue command or having complementary material search and rescue functions), the corresponding mask value is 0; if there is no social connection, the mask value is set to an engineering truncation negative infinity constant. This mask is directly injected as a structural bias into the scaled dot product self-attention score. Utilizing the zeroing property of negative infinity in the exponential Softmax operation, topological interference noise with no social background to the current rescue mission is forcibly cut off, forcing features to flow only along the "implicit social bridge" induced by social homogeneity. Thus, even when the connection between two search and rescue teams is broken due to obstruction, the potential contact patterns between them can still be accurately discovered by crossing the fault line.

[0140] Secondly, post-disaster communication exhibits a typical pattern of "long-term disconnection followed by short-term, sudden encounters." During the extended period of no contact, if a traditional spatiotemporal decoupling model is used, the temporal module will blindly decay historical memory at a fixed rate, causing the sparse, critical contact features painstakingly established beforehand to be completely diluted and forgotten during the window of time, leading to misjudgments in link prediction. This invention directly intercepts the diagonal elements of the spatial social self-attention weight matrix (i.e., the social spatial interaction activity of the node at the current moment), and uses parameters... After gating adjustment, it is used as a dynamic adjustment factor for the continuous-time transition matrix to construct a spatiotemporal fusion state transition matrix. Discretization is performed using the zero-order hold (ZOH) rule to achieve selective linear recursive updates of the hidden states. That is, when a drone or search and rescue vehicle suddenly encounters and interacts in the disaster area, the memory retention coefficient of the system matrix is ​​dynamically amplified to approach 1, entering a "selective latching state," ensuring that even if there is a disconnection period of several days thereafter, this crucial encounter feature will never be forgotten; while when the network is in a continuous and lengthy period without any encounters, the system transition matrix adaptively accelerates its inherent decay, quickly filtering out temporally redundant white noise.

[0141] After undergoing a complete time window evolution, the link inference unit extracts the final high-order spatiotemporal coupling representation of the target node pair to be predicted and performs channel stitching. Then, through a multilayer perceptron decoder and a Sigmoid activation function mapping, it outputs the dynamic probability of the two devices making contact and connecting in the next time step. During the training phase, to address the severe class imbalance caused by the extremely low frequency of post-disaster encounters, a 1:1 dynamic negative sampling strategy was implemented to optimize the binary cross-entropy loss, thereby calculating a more accurate contact edge probability.

[0142] The embodiments of the present invention provide a unified social-spatiotemporal modeling method, characterized by including: local structure modeling: using spectral graph decomposition technology to capture structural evolution patterns under decentralized constraints; social semantic injection: eliminating prediction blind spots caused by structural discontinuities through a social mask self-attention mechanism; and decoupled evolutionary tracing: using selective state-space equations to achieve deep entanglement of spatiotemporal characteristics and adaptively filtering redundant information.

[0143] Compared with the prior art, the beneficial effects of the present invention include:

[0144] (1) This application introduces a symmetric normalized Laplace matrix with numerical protection factor and self-loop mechanism to extract local spectral features, enabling decentralized opportunistic networks to perform smooth and robust local structure modeling without relying on any global topological priors. This invention addresses the local observation constraints of opportunistic network topology by ensuring the feature flow path of isolated nodes with zero degree through self-loop augmentation, and completely eliminates computational collapse when the network is extremely sparse by introducing a minimal numerical protection factor in the degree matrix inversion and square root operation. This micro-design successfully maps discrete and fragmented topologies into numerically stable low-frequency structural anchors, significantly improving the structural characterization capability and numerical stability in decentralized environments compared to traditional spatial models that highly rely on dense, complete global topologies.

[0145] (2) This application features a sparsity-aware adaptive feature imputation mechanism (SAIM). The model adaptively activates a learnable missing token based on the real-time feature missing rate of the entire network, thereby completely eliminating semantic noise interference under incomplete attribute conditions. This invention quantifies the social feature missing rate of all network nodes online through an indicator function and sets a perceptual gating adaptive activation of the "missing token" at the feature mapping front end. In end-to-end gradient backpropagation, the missing token is used to automatically fit the latent implicit semantic distribution behind the missing nodes, completely eliminating the random multimodal noise introduced into the network by traditional zero-filling or static placeholder strategies, and greatly enhancing the robustness of the model in heterogeneous feature fusion under harsh data environments.

[0146] (3) This application constructs a spatial structure bias injection mechanism (SASInjector) based on social homogeneity masks. By cutting off the noise propagation of irrelevant nodes, it crosses physical topological faults and eliminates the prediction blind spots caused by structural fractures. This invention utilizes long-term relatively stable social attributes to reconstruct global homogeneity relationships and transforms them into a spatial topological mask matrix (the mask value of positions with common social attributes is zero, and the mask value of unrelated positions is negative infinity). When calculating the self-attention weights, this mask is seamlessly embedded as a structural bias. By utilizing the zero-exponential property of the negative infinity symbol in the Softmax operation, features are forced to flow only along the "implicit social bridge" induced by social homogeneity. This design utilizes the compensating effect of social semantics in the spatial dimension to solve the problem of feature fractures and topological prediction blind spots when relying solely on explicit connections for message passing.

[0147] (4) This application innovatively proposes a decoupled spatiotemporal coupling evolution architecture for spatial self-attention weight modulation of temporal memory, which solves the memory decay problem caused by spatiotemporal fragmentation and long-term lack of contact in traditional models from the physical computation level. This invention breaks the traditional serial-parallel decoupling paradigm of "first spatial convolution, then temporal loop", directly extracts the diagonal elements (node ​​self-attention) of the spatial self-attention weight matrix, and uses them as a dynamic regulator of temporal memory coefficients after nonlinear projection. This regulator is directly applied to the basic continuous state transition matrix A of the selective state space model (Mamba) to construct a highly entangled "spatiotemporal fusion state transition matrix". This design realizes the integrated fusion of spatiotemporal multidimensional features, enabling the system to adaptively adjust the temporal memory decay rate according to the strength of spatial social interaction. When strong social interaction occurs, memory decay is suppressed to lock key events, and information decay is accelerated during the no-contact window period. Thus, while ensuring that the temporal logic does not shift, it successfully crosses and filters out contact gaps that last for several weeks.

[0148] (5) This application establishes a joint optimization and boundary equilibrium mechanism based on a dynamic negative sampling strategy using binary cross-entropy loss, ensuring efficient and robust convergence of the model under conditions of extreme topological sparsity and severe class imbalance. Addressing the hard constraint of extreme class imbalance where positive samples in the opportunistic network are far fewer than untouched negative samples, this invention activates dynamic negative sampling during the training phase. For each real-existing positive sample with a contact edge, a strictly equal number of samples are randomly selected from all non-contact node pairs in the entire network to form a negative sampling set. This boundary equilibrium mechanism effectively alleviates gradient skew during the network-wide optimization process, preventing the algorithm from falling into trivial solutions where all predictions are negative, and ensuring that the model can learn deep social-spatial mapping relationships across spatiotemporal discontinuities with high accuracy and efficiency.

[0149] (6) Since this application does not rely on global topology flow priors, specific workload features, or hard-coded dataset rules, it has excellent parameter adaptability and engineering deployment flexibility under the hard constraints of shared topology observations. The selective state-space equations and zero-order preserved discretization adopted by the system only require linear computational complexity, which not only significantly reduces the optimization difficulty and computational overhead when fusing multi-source heterogeneous features, but also gives it excellent scenario generalization. This invention can be incrementally used as a node-level reserved coordination submodule and can be widely deployed in ubiquitous distributed Internet of Things scenarios such as decentralized self-organizing networks, vehicle-to-everything (V2X) roadside unit collaborative systems, post-disaster emergency self-organizing gateways, and smart pastoral mobile network monitoring.

[0150] Finally, it should be noted that although this specification contains many specific implementation details, these should not be construed as limiting the scope of any invention or the scope of the claims, but rather are primarily used to describe the features of specific embodiments of a particular invention. Certain features described in the various embodiments of this specification may also be implemented in combination in a single embodiment. On the other hand, various features described in a single embodiment may also be implemented separately in various embodiments or in any suitable sub-combination. Furthermore, while features may function in certain combinations as described above and even initially claimed in this way, one or more features from a claimed combination may be removed from that combination in some cases, and a claimed combination may refer to a sub-combination or a variation of a sub-combination.

[0151] Similarly, although the operations are depicted in a specific order in the accompanying drawings, this should not be construed as requiring these operations to be performed in the specific order shown or sequentially, or requiring all illustrated operations to be performed to achieve the desired result. In some cases, multitasking and parallel processing may be advantageous. Furthermore, the separation of various system modules and components in the above embodiments should not be construed as requiring such separation in all embodiments, and it should be understood that the described program components and systems can generally be integrated together in a single software product or packaged into multiple software products.

[0152] Thus, specific embodiments of the subject matter have been described. Other embodiments are within the scope of the appended claims. In some cases, the actions recited in the claims may be performed in a different order and still achieve the desired result. Furthermore, the processes depicted in the drawings are not necessarily shown in a specific order or sequence to achieve the desired result. In some implementations, multitasking and parallel processing may be advantageous.

[0153] The above description is merely a preferred embodiment of this disclosure and is not intended to limit this disclosure. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A dynamic link prediction method for opportunity networks based on a decoupled social-spatiotemporal selective state-space model, characterized in that, The method is applied to nodes in opportunistic networks that lack global topological priors and have incomplete heterogeneous social attributes, and includes the following steps: The raw contact log stream generated by the mobile node of the receive opportunity network is divided into discrete time steps according to a preset resolution and then processed within a sliding time window. The initial discrete adjacency matrix A(t) is constructed by extracting local physical interaction records. A self-loop mechanism is introduced into the initial off-connection matrix A(t), and an identity matrix is ​​added on the diagonal. Generate augmented local adjacency matrix Where N represents the total number of all devices; Based on the augmented local adjacency matrix Calculate the degree matrix D(t), and introduce a numerical protection factor into the diagonal elements of the degree matrix D(t). A symmetric normalized Laplace matrix is ​​constructed using a degree matrix with a numerical protection factor. ; for the Laplace matrix Perform real symmetric eigenvalue decomposition, and after non-negativity truncation, select the eigenvectors corresponding to the k smallest eigenvalues ​​as stable low-frequency local spectral features. ; By defining an indicator function to scan the original social attributes of the input node by node, the social feature missing rate of all nodes in the network at the current time step is quantitatively calculated. According to the missing rate Adaptive gating activation of learnable missing tokens. The latent semantic distribution behind the missing behavior is automatically learned through gradient backpropagation and implicit completion is performed. Temporary state features are spliced ​​to generate an aligned initial representation matrix X(t). The initial representation matrix X(t) is projected through three independent linear mapping layers to generate a query matrix Q(t), a key matrix K(t), and a value matrix V(t). Global homogeneous relationships are then constructed using long-term static social attributes and transformed into a spatial topological mask matrix. The spatial topology mask matrix As a spatial structure bias, it is seamlessly embedded in the scaling dot product self-attention score calculation in an additive form. Utilizing the zero-exponential property of the negative infinity sign in the Softmax operation, it cuts off noise propagation from nodes without social connections, outputting spatially enhanced features. In this context, SA represents that the feature has incorporated socio-semantic priors; Capture the diagonal elements of the self-attention weight matrix The nonlinear projection of this matrix is ​​used as a dynamic adjustment factor for the time memory coefficient, and then superimposed on the basic continuous state transition matrix A of the selective state-space model to construct the spatiotemporal fusion state transition matrix. Introduce a zero-order preserve rule and dynamically adaptive step size based on input dependence. Discretization transformation is performed to realize the hidden state of the node. The first-order linear selective recursive update utilizes the intensity of spatial social interaction to adaptively filter lengthy contactless periods and lock key contact events. It is the social self-attention weight matrix at the t-th sliding time step; After undergoing a complete time window recursive evolution, the final coupled hidden state representations of the target prediction node pairs are extracted and concatenated along the channel dimension. The contact connection probabilities for future time steps are then inferred using a multilayer perceptron decoder and a sigmoid activation function. Where u and v represent any two independent individuals in the set of network nodes V. Indicates the current moment. This represents the moment in the future that is immediately following historical observations and yet to be predicted.

2. The method according to claim 1, characterized in that, The Laplace matrix For microscopic construction, the numerical protection factor This is to prevent numerical collapse caused by performing inversion or square root operations on the degree matrix when the degree of the opportunistic network node is zero in the current time slice.

3. The method according to claim 1, characterized in that, The adaptive feature interpolation is performed using the following steps: When the original social attribute vector of the target node is determined by the indicator function When the vector is all zeros, the learnable "missing tokens" are forcibly padded at the front end of the feature map; When the original social attribute vector of the target node is determined to be non-empty, the social features are mapped to a dense and continuous semantic embedding space through a linear projection layer to eliminate the random multimodal noise introduced by the zero-vector hard-filled directed graph network.

4. The method according to claim 1, characterized in that, In the feature mapping front end of the heterogeneous encoder, for any node i, a feature based on... Adaptive interpolation formula for sociosemantic embedding with dual constraints of local norm: ； in, The learnable weight matrix for the linear projection layer of social features. For the learnable bias vector of the linear projection layer of social features. Let be the socio-semantic embedding representation of node i at time step t after adaptive interpolation, where t is the current sliding historical time step. , where i is the index of the target node currently being scanned by the interpolation algorithm.

5. The method according to claim 4, characterized in that, The adaptive feature interpolation is performed using the following steps: When the original social attribute vector of the target node is determined by the indicator function When the vector is all zero, the learnable missing tokens are forcibly padded at the front end of the feature map; When the original social attribute vector of the target node is not empty, the social features are mapped to a dense and continuous semantic embedding space through a linear projection layer to eliminate the random multimodal noise introduced by the zero-vector hard-filled directed graph network.

6. The method according to claim 1, characterized in that, The spatial topology mask matrix This includes micro-element setting rules, which cause the spatial attention score between nodes without social background to be compressed to zero after exponential normalization.

7. The method according to claim 6, characterized in that, Spatial topological mask matrix Its microscopic elements are mathematically constructed according to a strict step-asymmetric definition: ； in, It is a matrix The specific element value in the i-th row and j-th column, Represents the dimension of the matrix. Represents the topological mask matrix It is a two-dimensional square matrix consisting of N rows and N columns of real numbers.

8. The method according to claim 1, characterized in that, The construction of the spatiotemporal fusion state transition matrix is ​​used to balance the intensity of dynamic spatial social interaction with the decay rate of inherent temporal historical memory.

9. The method according to claim 1, characterized in that, The dynamic negative sampling strategy performs the following steps: For each positive sample edge in the current training batch that has had an actual topological contact A negative sampling set is formed by randomly selecting an equal number of samples from node pairs that have no actual contact records across the entire network. This ensures that the training target set for calculating the binary cross-entropy loss function maintains a 1:1 balance in the positive and negative class distributions.

10. The method according to claim 1, characterized in that, The scale of the sliding time window in the evolutionary sequence is fixed at [value]. Each discrete time step has a sampling time resolution between 270 seconds and 3600 seconds, depending on the dynamic change frequency of the decentralized application scenario.

Images