Dynamic link prediction method and system based on snapshot redivision and incremental learning, and storage medium

Abstract

The invention discloses a dynamic link prediction method and system based on snapshot redivision and incremental learning and a storage medium, and the method comprises the steps: constructing initial snapshots, dividing the initial snapshots into time span adaptive sub-snapshots based on a network evolution rate, flattening the sub-snapshots into a new snapshot sequence, enabling a current snapshot to inherit a topological structure of a preorder snapshot, and enabling the current snapshot to be a new snapshot sequence; dividing each snapshot node set into a newly added node set and an inherited node set; performing biased random walk by taking a newly added node set in the snapshot as a start to generate a new walk sequence, and obtaining an initial node representation matrix according to the new walk sequence; performing multi-layer topological feature aggregation on the newly added nodes through an incremental multi-layer graph attention network to obtain a topological feature matrix; the method comprises the following steps of: standardizing time spans, splicing the standardized time spans into topological features, inputting a time perception long-short-term memory network to obtain node representation with enhanced time attributes, establishing nonlinear correlation among nodes through a multi-layer perceptron, generating a probability matrix for forming edges among the nodes, and training a model by adopting a binary cross entropy loss function.

Description

TECHNICAL FIELD

[0001] The present application particularly relates to a dynamic link prediction method system based on snapshot re-partitioning and incremental learning and a storage medium. BACKGROUND

[0002] Networks in the real world, such as social networks, biological networks, communication networks, etc., generally have dynamic evolution characteristics, and the connection relationship between network nodes will change over time. Dynamic link prediction, as a core task of dynamic network analysis, aims to predict the possible node connections in the future by mining the historical evolution rules of the network, and provides key support for network behavior understanding and decision optimization.

[0003] The current mainstream dynamic link prediction method is mostly based on the discrete time dynamic graph (DTDG) framework, which discretizes the continuous evolution process of the network into a series of network snapshots by setting a fixed time span τ, and then makes predictions based on the historical snapshot rules. However, this kind of method has significant limitations: for example, the evolution of real networks often presents an alternating pattern of "pulsed mutation" and "gradual evolution", and the fixed time span is difficult to match the evolution rate at different stages. Intensive events in the rapid change period will be compressed into coarse-grained snapshots, resulting in the loss of key information; and redundant snapshots will be generated in the slow change period, increasing the computational cost. In addition, adjacent snapshots in DTDG are independent of each other, ignoring the time continuity of the network topology, and the model cannot capture the gradual evolution rules across multiple time windows, weakening the temporal coherence of the prediction results. Existing graph representation learning methods, such as random walk and graph neural network, require repeated computation of the full amount of nodes and edges for each snapshot when processing dynamic snapshots, and even if the network only changes locally, all node representations need to be relearned, resulting in a large amount of redundant computation and making it difficult to adapt to large-scale dynamic networks. Some current methods improve the graph neural network to enhance the temporal modeling capability, but do not solve the problem of evolution rate mismatch caused by fixed time span; some methods introduce incremental learning to reduce redundant computation, but lack effective modeling of the temporal continuity between snapshots; and some methods optimize snapshot partitioning by dynamically adjusting the time window, but do not deeply integrate with subsequent graph representation learning and temporal modeling modules, limiting the overall prediction performance.

[0004] Therefore, there is an urgent need for a dynamic link prediction method to solve the above technical problems. SUMMARY

[0005] To solve the problems mentioned in the background, the purpose of the present application is to provide a dynamic link prediction method based on snapshot re-partitioning and incremental learning, comprising: S1, data preprocessing and initial snapshot construction: obtaining dynamic network original evolution data, discretizing into an initial snapshot sequence according to an initial fixed time span; S2, adaptive snapshot re-partitioning: based on the network evolution rate , each initial snapshot is divided into time span adaptive sub-snapshot and flattened into a new snapshot sequence; S3, inter-snapshot state inheritance: make the current snapshot inherit the topology structure of the previous snapshot, divide the snapshot node set into a new node set and an inherited node set ; wherein is the node set of the original snapshot before the inheritance operation is performed, is the node set inherited from the previous snapshot; S4, incremental node representation generation: perform biased random walk with the new node set in the current snapshot as the starting node to generate a new walk sequence, train the new walk sequence through the Word2Vec method to obtain the new node representation , the historical representation of the inherited node is denoted as , the new node representation and the inherited node representation together constitute the initial node representation matrix of the current snapshot ; S5, incremental topology information learning: through the incremental multi-layer graph attention network, only the new nodes are subjected to multi-layer topology feature aggregation, and the historical topology features of the inherited nodes are reused to obtain the topology feature matrix; S6, time attribute learning: through the time-aware long short-term memory network, set the standard time span of the current snapshot, splice the standard time span as an additional time feature into the topology feature, and introduce a time decay factor to dynamically decay the memory state of the previous time step to process non-equidistant snapshots, and obtain the time attribute enhanced node representation, wherein represents an adjustable hyperparameter that controls the strength of time decay; S7, link prediction and result training: through the multi-layer perceptron, establish the non-linear association between nodes to generate the probability matrix of the edges formed between nodes, and use the binary cross-entropy loss function to train the difference between the predicted results and the true labels.

[0006] Preferably, the adaptive snapshot re-partitioning step based on the network evolution rate includes: In the above formula, denotes the cardinality of the edge set , is the floor operator, denotes the original snapshot, denotes the number of sub-snapshots, denotes the time span of each sub-snapshot, denotes the th sub-snapshot after re-partitioning, denotes the node set and edge set observed in the corresponding time window, respectively, is the corresponding adjacency matrix; wherein the flattening is a new snapshot sequence process performed in the order of t as the primary sequence and k as the secondary sequence; the flattened snapshot sequence is denoted as wherein, represents the total number of flattened snapshots, represents the time span of the flattened snapshot .

[0007] Preferably, the process of inheriting the topology of the previous snapshot by the current snapshot comprises:

[0008] In the above formula, denotes any snapshot, denotes the snapshot after inheritance, wherein the snapshot located at the first position of the sequence does not perform the inheritance operation.

[0009] Preferably, the transition probability calculation formula of the biased random walk in step S4 is: In the above formula, denotes the clustering coefficient of node , denotes the closeness centrality of node , denotes the betweenness centrality of node , denotes the degree of node , denotes the one-hop neighbor set of node , denotes an adjustable parameter.

[0010] Preferably, the feature update formula of the incremental multi-layer graph attention network in step S5 is: In the above formula, denotes the node In the incremental multi-layer graph attention network layer feature representation, and is a nonlinear activation function, represents the one-hop neighbor set of the node , and is a weight parameter to be learned.

[0011] Preferably, the standard time span of the current snapshot is set, and the standard time span is spliced into the topology feature in the splicing process as an additional time feature: In the above formula, represents the time span of the current snapshot , and respectively represent the mean and standard deviation of all snapshot time spans in the training set.

[0012] Preferably, the binary cross-entropy loss function is:

[0013] In the above formula, represents the total number of nodes, represents the corresponding true label, represents the predicted link probability.

[0014] The application also provides a system for dynamic link prediction, comprising: a data preprocessing module: for obtaining original evolution data of a dynamic network and constructing an initial snapshot sequence; a snapshot processing module: for realizing adaptive snapshot redivision and snapshot state inheritance; an incremental biased random walk module: for generating initial node representation; an incremental multi-layer graph attention network module: for multi-layer topology feature aggregation of new nodes, inheritance of node reuse of historical topology features, and obtaining of a topology feature matrix; a time series modeling module: for modeling time series dependence through a time-aware long short-term memory network, and obtaining of a time attribute enhanced node representation; a prediction and training module: for outputting a connection probability through a multi-layer perceptron, and training a model through a binary cross-entropy loss function.

[0015] Preferably, the calculation process of the incremental multi-layer graph attention network module is: inputting the initial node representation matrix generated by the incremental random walk , the node set , and the node feature representation of the previous snapshot ; Calculate and output the node representation matrix after the current snapshot has completed topological feature learning. .

[0016] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described dynamic link prediction method based on snapshot repartitioning and incremental learning.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: I. Adapting to Non-uniform Evolution: Through adaptive snapshot repartitioning, the time span is dynamically adjusted based on the network evolution rate. Fine-grained snapshots are generated during periods of rapid change to retain key information, while coarse-grained snapshots are generated during periods of slow change to reduce redundancy, effectively matching the non-uniform evolution characteristics of real-world networks.

[0018] 2. Maintaining temporal continuity: The state inheritance mechanism between snapshots forces the current snapshot to inherit the topology of the previous snapshot, avoiding the separation between adjacent snapshots, providing the model with a continuous evolution trajectory, and improving the temporal coherence of the prediction results.

[0019] Third, improve computational efficiency: The incremental learning mechanism runs through the node representation generation and topology information learning process. It only performs calculations on newly added nodes, inherits nodes to reuse historical results, reduces redundant calculations, lowers the processing cost of large-scale dynamic networks, and improves processing speed.

[0020] IV. Enhanced Prediction Accuracy: The biased random walk that integrates the influence of node structure improves the recognizability of node representations; the incremental multi-layer graph attention network captures multi-order topological dependencies; a time-aware mechanism is introduced; the time-aware long short-term memory network adapts to the time series modeling of non-equidistant sequences; and the multi-module collaboration improves the accuracy and generalization ability of link prediction. Attached Figure Description

[0021] For ease of explanation, the present invention will be described in detail below with reference to specific embodiments and accompanying drawings.

[0022] Figure 1 The diagram shows the overall model of the dynamic link prediction method based on snapshot repartitioning and incremental learning. Figure 2 A scatter plot of the number of evolutionary events for the initial snapshots of each dataset; Figure 3 Line plots showing the AUC and AP of SRIL as a function of parameter a on different datasets; Figure 4 This is a diagram of a dynamic link prediction system based on snapshot repartitioning and incremental learning. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the invention is described below with reference to specific embodiments shown in the accompanying drawings. However, it should be understood that these descriptions are merely exemplary and not intended to limit the scope of the invention. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.

[0024] It should also be noted that, in order to avoid obscuring the invention with unnecessary details, only the structures and / or processing steps closely related to the solution according to the invention are shown in the accompanying drawings, while other details that are not closely related to the invention are omitted.

[0025] Specific Implementation Method 1: This embodiment provides only a preferred implementation method. This invention provides a Dynamic Link Prediction (SRIL) method based on snapshot re-partitioning and incremental learning. The SRIL method abandons the traditional fixed time span strategy and adopts a data-driven adaptive mechanism. It dynamically adjusts the time span of snapshots by monitoring the network change rate and performs secondary partitioning of snapshots. For example, during a period of intense public opinion on a social network, the model can automatically shorten the time span to capture a more refined network evolution process. To address the issue of inconsistent snapshot time spans caused by adaptive partitioning, improvements are made to the Long Short-Term Memory (LSTM) network by introducing the time span as an additional input, thereby enhancing the model's ability to model non-equidistant sequences. Regarding evolutionary continuity modeling, SRIL establishes a state inheritance mechanism between snapshots, integrating the state information of previous snapshots into the current snapshot, effectively improving the model's ability to express cross-snapshot dependencies and maintaining the temporal coherence of network evolution. However, since the state inheritance mechanism may lead to the reuse of historical information, it increases computational redundancy in the node representation learning process. Therefore, the SRIL method also introduces an incremental learning mechanism to improve two graph representation learning algorithms: Graph Random Walk and Graph Attention Network (GAT). SRIL only focuses on the incremental parts of the network that have changed, effectively avoiding redundant computation. Specifically, for example... Figure 1As shown, the SRIL method mainly includes the following parts: a) Adaptive snapshot repartitioning and state inheritance between snapshots: SRIL first adaptively repartitions the original snapshot sequence and performs state inheritance operations between snapshots, thereby modeling the dynamic evolution process of the network more finely; b) Node representation generation: State inheritance divides the nodes in each snapshot into new nodes and inherited nodes. For new nodes, SRIL uses incremental biased random walk and word2vec methods to generate the initial representation of the node, while inherited nodes can directly reuse their historical representations to avoid redundant calculations; c) Topology information learning: SRIL learns the topology information of new nodes in the current snapshot layer by layer through an incremental multi-layer graph attention network, while inherited nodes still reuse their historical representations; d) Temporal information learning and link prediction: After the topology information of all nodes is learned, SRIL uses a time-aware long short-term memory network (T-LSTM) to model the temporal dependency features of nodes in the snapshot sequence. The introduction of topology information and temporal dependency information effectively enriches the semantic information and expressive power of the node representation. The model predicts the probability of forming links between node pairs based on the updated node representations and outputs a connection probability matrix.

[0026] S1. Data preprocessing and initial snapshot construction: Obtain the original evolution data of the dynamic network, which includes a set of nodes, a set of edges, and a timestamp corresponding to each edge. Discretize the original evolution data into an initial snapshot sequence according to a preset initial fixed time span.

[0027] S2. Adaptive Snapshot Re-partitioning: Adaptively adjusts the temporal resolution based on the network evolution rate to optimize snapshot partitioning and define the minimum network evolution rate. The minimum cardinality of all snapshot edge sets is obtained through this method, which yields the original snapshot. Reclassified into A smaller sub-snapshot with a shorter time span, based on the network evolution rate. The adaptive snapshot repartitioning step divides each initial snapshot into sub-snapshots with an adaptive time span and flattens them into a new snapshot sequence. This adaptive partitioning method effectively adapts to the non-uniform characteristics of network evolution rate by dynamically adjusting the time resolution. S3. Snapshot State Inheritance: To more accurately reflect the temporal continuity of network evolution, this invention proposes an inter-snapshot state inheritance method. This method allows the current snapshot to inherit the topology of its predecessor snapshot, allowing any snapshot to inherit its topology. After the inheritance operation is completed, it is represented as The snapshot at the beginning of the sequence has no predecessor snapshot, so no inheritance operation is performed. After progressive topology overlay, each snapshot... Node set in It can be decomposed into two complementary subsets, with the addition of a node set. With inheritance node set .here, That is, the original snapshot before the inheritance operation was performed. node set ,and These are nodes that inherit from the previous snapshot, represented as In the process of learning node representations, inherited nodes can directly reuse historical results, and only the representation learning needs to be performed on newly added nodes. This method ensures that there is a stable intersection between the node set and edge set between adjacent snapshots, effectively solves the problem of context information fragmentation, provides continuous structural evolution trajectory information for the prediction model, and only calculates for newly added nodes, reducing redundancy.

[0028] S4. Incremental Node Representation Generation: This invention also proposes an incremental biased random walk model. It integrates a random walk mechanism with a node influence bias strategy and introduces an incremental update mechanism. The incremental update strategy uses only newly added nodes in the snapshot as the starting point for the random walk, avoiding the high computational overhead of repeatedly traversing the entire network in traditional methods. Simultaneously, it introduces a node influence bias mechanism on top of the traditional random walk, where, in each step of the walk, the node influence bias is adjusted from the current node... Migrate to its neighboring node transition probability Instead of following a uniform distribution, the path is adjusted based on the influence of nodes, favoring jumps to neighboring nodes of greater structural importance. The introduction of node influence makes the traversal path more focused on key regions of the network, thereby improving the expressive power and structural discernibility of node representations. Specifically, for the newly added node set in the current snapshot… Starting from these nodes, the biased random walk described above is performed to generate new walk sequences. Subsequently, the Word2Vec

[10] method is used to train these newly generated sequences to obtain low-dimensional embedding representations of the newly added nodes. In contrast, for inherited nodes in a snapshot, their historical representation is directly reused and denoted as... Ultimately, the newly added node represents Representation of inheritance nodes Together they constitute the current snapshot The initial node representation matrix ; S5. Incremental Topological Information Learning: In traditional graph neural networks, the receptive field is typically expanded by increasing the number of model layers, enabling nodes to aggregate information from more distant neighbors. However, blindly stacking network layers can easily lead to ineffective diffusion and redundant information transmission of node features in multiple rounds of propagation, not only reducing computational efficiency but also potentially weakening the discriminative ability of node representations. Therefore, this invention provides an Incremental Multi-layer Graph Attention Network (IMGAT) module for learning the topological structural features of nodes. This module fuses the feature information of nodes with their neighbors of different orders, performs multi-layer topological feature aggregation on newly added nodes, inherits and reuses historical topological features of nodes, captures the topological dependencies of nodes in multi-order neighborhoods, and obtains a topological feature matrix, thereby effectively improving the discriminative ability of node representations. This invention also provides a progressively decreasing feature dimension. As the number of IMGAT layers increases, corresponding to the aggregation of higher-order neighbor information, the feature dimension of the node representation gradually decreases, which can encourage the model to focus more on high-value information from nearby nodes and avoid ineffective diffusion. S6. Time Attribute Learning: This invention also provides a Time-aware Long Short-Term Memory (LSTM) network (T-LSTM), which improves LSTM by introducing a time-aware mechanism, enabling it to adapt to snapshot sequences with different time spans. Specifically, T-LSTM first standardizes the time span of the current snapshot and concatenates the standard time span as an additional time feature into the node's representation vector to explicitly enhance the model's time-aware capability. The concatenated input not only contains structural information but also explicitly introduces the snapshot's time span, helping the model capture the evolutionary patterns under non-uniform time intervals. Furthermore, T-LSTM introduces a time decay factor into the forgetting gate of traditional LSTM. The memory state of the previous time step is dynamically decayed, and non-uniformly spaced snapshots are processed to obtain node representations with enhanced time attributes. This allows for dynamic control of the retention rate of old information based on the time span. The adjustable hyperparameter represents the intensity of time decay control; the overall calculation process of T-LSTM is shown in equations (11) to (17).

[0029] in, All are learnable weight matrices. All are learnable bias vectors, with signs... This represents element-wise product, also known as the Hadamard product. For the first Memory state information at each time step To complete the node representation after learning the temporal attributes; using the time decay factor. Memory state of the previous time step Dynamic decay is employed, enabling the model to automatically adjust the influence of historical information on the current state based on different time intervals between snapshots, thereby achieving accurate modeling of non-equal interval snapshot sequences.

[0030] S7. Link Prediction and Result Training: Preferably, this invention transforms the link prediction task into a binary classification problem and employs a multilayer perceptron (MLP) as the classifier. The node representations learned by the aforementioned modules... The data is input into an MLP to model the nonlinear relationships between node pairs, and finally generates a probability matrix of edges between nodes in the output layer. probability matrix The calculation process is as follows: in, It is a non-linear activation function. Represents a node and nodes There is a probability that there is an edge between them. A value close to 1 indicates a high probability that an edge exists between the two nodes. Furthermore, this invention uses a binary cross-entropy loss function to train the model on the difference between the predicted results and the true labels.

[0031] Specific Implementation Method Two: Content not mentioned in this implementation method is the same as in Specific Implementation Method One. The above is based on network evolution rate. The adaptive snapshot repartitioning step, which divides each initial snapshot into sub-snapshots with an adaptive time span and flattens them into a new snapshot sequence, includes: In the above formula, Representing edge set The base number, This is the floor operator. Indicates the original snapshot. Indicates the number of sub-snapshots. This indicates the time span of each sub-snapshot. Represents the first after repartitioning Each snapshot corresponds to a time window. The network state within the system effectively captures fine-grained events during periods of rapid, impulsive change. These represent the set of nodes and the set of edges observed within the corresponding time window, respectively. This is the corresponding adjacency matrix; The process of flattening the snapshot sequence into a new snapshot sequence is performed with t as the primary sequence and k as the secondary sequence; the flattened snapshot sequence is denoted as... ,in, This represents the total number of snapshots after flattening. Snapshot of the flattened representation The time span.

[0032] Specific Implementation Method Three: Content not mentioned in this implementation method is the same as in Specific Implementation Method One or Two. The process of the current snapshot inheriting the topology of the previous snapshot includes:

[0033] In the above formula, Represents any snapshot. This represents the snapshot after inheritance, where the snapshot at the beginning of the sequence does not undergo inheritance.

[0034] Specific Implementation Method Four: Content not mentioned in this implementation method is the same as in Specific Implementation Methods One, Two, or Three. The formula for calculating the transition probability of the biased random walk is as follows: In the above formula, Represents a node Clustering coefficient, Represents a node Proximity centrality Represents a node betweenness centrality, Represents a node The degree, Represents a node A jump neighbor set, This indicates an adjustable parameter.

[0035] Specific Implementation Method Five: Content not mentioned in this implementation method is the same as in Specific Implementation Methods One, Two, Three, or Four. The feature update formula for the incremental multilayer graph attention network is: In the above formula, Represents a node In incremental multilayer graph attention network Layer feature representation, and It is a non-linear activation function. Represents a node A jump neighbor set, and These are the weight parameters to be learned.

[0036] Specific Implementation Method Six: Content not mentioned in this implementation method is the same as in Specific Implementation Methods One, Two, Three, Four, or Five. The process of setting the standard time span of the current snapshot and stitching it into the topological features using the standard time span as an additional time feature is as follows: In the above formula, Indicates the current snapshot Time span, and represents the mean and standard deviation of all snapshot time spans in the training set, respectively.

[0037] Specific Implementation Method Seven: Content not mentioned in this implementation method is the same as in Specific Implementation Methods One, Two, Three, Four, Five, or Six. This invention is a training model that uses a binary cross-entropy loss function to measure the difference between the predicted result and the true label. The binary cross-entropy loss function is: In the above formula, Indicates the total number of nodes. This represents the corresponding real label, used to indicate the node pair. In snapshot Does an actual connection exist in it? This indicates the predicted link probability.

[0038] Specific Implementation Method Eight: This embodiment provides only a preferred implementation method. The present invention provides a system for dynamic link prediction to implement the above method, including: Data preprocessing module: used to acquire raw evolution data of dynamic networks and construct initial snapshot sequences; Snapshot processing module: used to implement adaptive snapshot repartitioning and state inheritance between snapshots; Incremental biased random walk module: used to generate initial node representations; Incremental multi-layer graph attention network module: used to perform multi-layer topological feature aggregation on newly added nodes, inherit the historical topological features of nodes, and obtain the topological feature matrix; Temporal modeling module: used to model temporal dependencies through time-aware long short-term memory networks to obtain node representations with enhanced temporal attributes; Prediction and training module: used to output connection probabilities through a multilayer perceptron and train the model using a binary cross-entropy loss function; the above system can be deployed on servers, cloud computing platforms or edge computing devices, and supports real-time output of dynamic network data.

[0039] Specific Implementation Method Nine: This embodiment provides only a preferred implementation method. Content not mentioned in this embodiment is the same as in Specific Implementation Method Eight. The calculation algorithm for the above-mentioned incremental multilayer graph attention network module is as follows:

[0040] In the algorithm, the second line represents the initial node representation matrix. As a node representation of layer 0 of IMGAT, denoted as L represents the maximum number of layers in IMGAT. Lines 4 to 6 perform topological feature learning layer by layer on the newly added nodes in the snapshot. Line 8 concatenates the features from all layers to obtain the final representation of the newly added node, denoted as . Line 9 merges the representation of the newly added node with the representation of the inherited node in the previous snapshot to obtain the complete node representation. .

[0041] Specific Implementation Method Ten: This invention also provides a specific implementation method. Experiments were conducted on five real-world network datasets, and the basic statistical information of each dataset is shown in Table 1. To more intuitively demonstrate the size differences of each snapshot under a fixed time span, as shown... Figure 2 As shown, a scatter plot is used to present the number of network evolution events contained in all snapshots. Different colors and spatial locations are used to distinguish different datasets. Each scatter point represents a snapshot, with the horizontal axis representing the snapshot number; smaller numbers indicate earlier times within the same dataset. The vertical axis represents the number of evolution events contained in the snapshot; larger values ​​indicate a more "dense" snapshot. The plot shows that the snapshot sizes of the bitcoinotc and wikipedia datasets are generally stable with small differences. The collegeMsg dataset shows significant differences in snapshot sizes in the early stages, then tends to stabilize. The PPI dataset shows similar snapshot sizes initially, but the differences gradually increase later. The mooc_actions dataset, however, exhibits significant fluctuations in snapshot size throughout the entire time range, demonstrating a clear non-uniform evolution characteristic.

[0042] Table 1. Statistical Information of Dataset

[0043] For the data presented in Table 1, collegeMsg represents message communication records between users in the University of California, Irvine's online student community; mooc_action represents interactive behaviors performed by users on the MOOC platform; bitcoinotc represents the trust relationship network between users in the Bitcoin over-the-counter trading platform; PPI represents the protein-protein interaction network, where edges represent experimentally verified biological interactions between two proteins; and wikipedia represents the collaborative network formed by users' editing and updating operations on Wikipedia pages. Based on the above explanation, the experimental datasets selected in this embodiment cover multiple application areas, including communication networks, social networks, financial transactions, bioinformatics networks, and collaborative editing networks, and are highly representative. These datasets contain both sparse and dense networks, with diverse data structures, which helps to comprehensively verify the performance and robustness of the proposed method in different scenarios.

[0044] This embodiment preferably uses two commonly used metrics for link prediction, AUC (Area Under the ROC Curve) and AP (Average Precision), to evaluate the model's prediction accuracy on the test set. Both reflect prediction performance from different perspectives; AUC emphasizes overall ranking quality, while AP emphasizes the initial accuracy of the prediction results.

[0045] AUC focuses on the model's overall ranking performance, while AP focuses more on the accuracy of high-confidence predictions at the top. By using both metrics simultaneously, a comprehensive evaluation of the model's performance in link prediction tasks can be achieved.

[0046] This invention conducted experiments on five real-world network datasets and compared them with four baseline methods: Roland GNN, DGCN, GC-LSTM, and EvolveGCN. Table 2 shows the experimental results of each method on different datasets. It can be seen that RSIL achieved the best performance on all datasets and evaluation metrics, fully validating its broad adaptability and superiority across various network architectures.

[0047] Table 2 Comparison of AUC and AP metrics for different methods on different datasets

[0048] In the mooc_actions dataset, SRIL achieved an AUC of 0.9907. In the PPI dataset, SRIL showed the most significant improvement, particularly in AP, exceeding the best-performing baseline method DGCN by 0.1555. This result fully validates the effectiveness of the adaptive snapshot repartitioning strategy in handling non-uniformly evolving networks and demonstrates SRIL's good generalization ability in sparse network environments. In the Wikipedia dataset, SRIL addressed the issue of baseline methods' difficulty in recognizing high-order neighbors, resulting in suboptimal prediction performance, by introducing an incremental multi-layer graph attention network, achieving an AUC improvement of 0.1091 compared to the best baseline method DGCN.

[0049] In addition, in incremental biased random walks, the parameters Used to adjust the weights of clustering coefficients and centrality indices in the transition probability. To explore... To assess the impact on model performance, this paper increases the step size from 0.1 to 0.9 by 0.2 and conducts experimental comparisons. Figure 3 Showing different The prediction performance of SRIL on five datasets is shown. The results indicate that on the bitcoinotc, collegeMsg, and PPI datasets, when... The model performs best when the clustering coefficient is set to 0.5, indicating that the impact of clustering coefficient and centrality on node importance is roughly equivalent in this type of network, suggesting a certain degree of node clustering structure. In the mooc_actions dataset, the model performs well... The optimal performance is achieved when the value is 0.7, and the prediction effect is similar to... The correlation coefficient is directly proportional, indicating that the clustering coefficient plays a more significant role in this network, and there is a strong clustering structure among the nodes. The dense network with many nodes clustered together is the main reason for its good prediction performance on this dataset. In the Wikipedia dataset, when The model performs best when the value is 0.1, and the prediction performance is similar to... The correlation is inversely proportional. This indicates that in this network, node centrality plays a dominant role, clustering is weak, and a significant long-tail distribution exists. In summary, The optimal value varies depending on the dataset and can reflect the dominance of different types of topological attributes in the network structure.

[0050] To evaluate the specific contribution of each module to SRIL performance, this embodiment also designed three sets of ablation experiments, removing one or more modules of the SRIL respectively, and comparing their prediction results with the complete model. The specific ablation models include: First: Del_SI: Removes the state inheritance mechanism between snapshots.

[0051] Second: Del_IL: Remove incremental biased random walks and incremental multilayer graph attention networks.

[0052] Third: Del_RS: Remove the adaptive snapshot repartitioning mechanism and T-LSTM module.

[0053] Table 3 below shows the comparison results of the AUC and AP metrics of the above model on five real network datasets.

[0054] Table 3 Comparison of AUC and AP metrics for ablation models on different datasets

[0055] As shown in the table, all three ablation models performed worse than the complete SRIL model across all metrics. Compared to Del_SI, SRIL improved AUC and AP by an average of 0.0197 and 0.0247, respectively, indicating that the snapshot state inheritance mechanism effectively captures the temporal continuity between snapshots. Compared to Del_IL, AUC and AP improved by an average of 0.0159 and 0.0207, respectively, demonstrating that incremental biased random walks and incremental multilayer graph attention networks help generate higher-quality node embedding representations. Compared to Del_RS, AUC and AP improved by an average of 0.0528 and 0.0556, validating the effectiveness of adaptive snapshot repartitioning in modeling the non-uniform evolution rhythm of the network, and the advantages of T-LSTM in handling non-uniformly spaced snapshot sequences. In summary, the ablation experiments fully demonstrate the important role of each module in the performance improvement of SRIL, especially the adaptive snapshot repartitioning and T-LSTM modules, which contribute the most significantly to the overall performance improvement and are key factors in the superior model performance.

[0056] This embodiment provides only a preferred implementation method. This implementation method is a storage medium that stores at least one instruction. The at least one instruction is loaded and executed by a processor to implement the above-mentioned dynamic link prediction method based on snapshot repartitioning and incremental learning. The storage medium described in this embodiment includes, but is not limited to, hard disks, USB flash drives, embedded system storage instruction sets, etc., and supports real-time data processing and parameter updates.

[0057] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A dynamic link prediction method based on snapshot repartitioning and incremental learning, characterized in that, include: S1. Data Preprocessing and Initial Snapshot Construction: Obtain the original evolution data of the dynamic network and discretize it into an initial snapshot sequence according to an initial fixed time span; S2, Adaptive Snapshot Re-partitioning: Based on Network Evolution Rate The adaptive snapshot repartitioning step divides each initial snapshot into sub-snapshots with an adaptive time span and flattens them into a new snapshot sequence. S3, Snapshot State Inheritance: Enables the current snapshot to inherit the topology of the previous snapshot, and sets the snapshot nodes together. Divided into new node sets With inheritance node set ;in The set of nodes from the original snapshot before the inheritance operation is performed. A set of nodes inherited from the previous snapshot; S4. Incremental node representation generation: using the newly added node set in the current snapshot. A biased random walk is performed on the starting node to generate a new walk sequence. The new walk sequence is then trained using the Word2Vec method to obtain the representation of the newly added node. The inherited node reuses the history representation as . The newly added node indicates Representation of inheritance nodes Together they constitute the current snapshot The initial node representation matrix ; S5. Incremental Topology Information Learning: Through an incremental multi-layer graph attention network, multi-layer topology features are aggregated only for newly added nodes, and historical topology features of inherited nodes are reused to obtain a topology feature matrix. S6. Time Attribute Learning: Using a time-aware long short-term memory network, a standard time span for the current snapshot is set. This standard time span is then used as an additional time feature and concatenated into the topological features. A time decay factor is also introduced. The memory state of the previous time step is dynamically decayed, and non-uniformly spaced snapshots are processed to obtain a node representation with enhanced temporal attributes. An adjustable hyperparameter representing the intensity of time decay; S7. Link Prediction and Result Training: Establish nonlinear relationships between nodes using a multilayer perceptron, and generate a probability matrix of edges between nodes. The binary cross-entropy loss function is used to train the model to analyze the difference between the predicted results and the true labels.

2. The dynamic link prediction method based on snapshot repartitioning and incremental learning according to claim 1, characterized in that, The network evolution rate-based The adaptive snapshot repartitioning step, which divides each initial snapshot into sub-snapshots with an adaptive time span and flattens them into a new snapshot sequence, includes: In the above formula, Representing edge set The base number, This is the floor operator. Indicates the original snapshot. Indicates the number of sub-snapshots. This indicates the time span of each sub-snapshot. Represents the first after repartitioning Individual snapshot, These represent the set of nodes and the set of edges observed within the corresponding time window, respectively. This is the corresponding adjacency matrix; The process of flattening the snapshot sequence into a new sequence is performed with t as the primary sequence and k as the secondary sequence; the flattened snapshot sequence is denoted as... ,in, This represents the total number of snapshots after flattening. Snapshot of the flattened representation The time span.

3. The dynamic link prediction method based on snapshot repartitioning and incremental learning according to claim 2, characterized in that, The process by which the current snapshot inherits the topology of the previous snapshot includes: In the above formula, Represents any snapshot. This represents the snapshot after inheritance, where the snapshot at the beginning of the sequence does not undergo inheritance.

4. The dynamic link prediction method based on snapshot repartitioning and incremental learning according to claim 3, characterized in that, The formula for calculating the transition probability of the partial random walk in step S4 is as follows: In the above formula, Represents a node Clustering coefficient, Represents a node Proximity centrality Represents a node betweenness centrality, Represents a node The degree, Represents a node A jump neighbor set, This indicates an adjustable parameter.

5. The dynamic link prediction method based on snapshot repartitioning and incremental learning according to claim 4, characterized in that, The feature update formula for the incremental multilayer graph attention network in step S5 is: In the above formula, Represents a node In incremental multilayer graph attention network Layer feature representation, and It is a non-linear activation function. Represents a node A jump neighbor set, and These are the weight parameters to be learned.

6. The dynamic link prediction method based on snapshot repartitioning and incremental learning according to claim 5, characterized in that, The process of setting the standard time span for the current snapshot and then stitching it into the topological features using the standard time span as an additional time feature is as follows: In the above formula, Indicates the current snapshot Time span, and represents the mean and standard deviation of all snapshot time spans in the training set, respectively.

7. The dynamic link prediction method based on snapshot repartitioning and incremental learning according to claim 6, characterized in that, The binary cross-entropy loss function is: In the above formula, Indicates the total number of nodes. This represents the corresponding real label. This indicates the predicted link probability.

8. A system for dynamic link prediction, characterized in that, include: Data preprocessing module: used to acquire raw evolution data of dynamic networks and construct initial snapshot sequences; Snapshot processing module: used to implement adaptive snapshot repartitioning and state inheritance between snapshots; Incremental biased random walk module: used to generate initial node representations; Incremental multi-layer graph attention network module: used to perform multi-layer topological feature aggregation on newly added nodes, inherit the historical topological features of nodes, and obtain the topological feature matrix; Temporal modeling module: used to model temporal dependencies through time-aware long short-term memory networks to obtain node representations with enhanced temporal attributes; Prediction and Training Module: Used to train the model by outputting connection probabilities through a multilayer perceptron and employing a binary cross-entropy loss function.

9. A system for dynamic link prediction according to claim 8, characterized in that, The computation process of the incremental multi-layer graph attention network module is as follows: the input is the initial node representation matrix generated by the incremental random walk. Node set and node feature representation of previous snapshots ; Calculate and output the node representation matrix after the current snapshot has completed topological feature learning. .

10. A computer-readable storage medium, characterized in that, The device contains a computer program that, when executed by a processor, implements the dynamic link prediction method based on snapshot repartitioning and incremental learning as described in any one of claims 1 to 7.

Images