Dynamic network core node prediction method and device, equipment and storage medium

Abstract

The application discloses a dynamic network core node prediction method and device, equipment and a storage medium, comprising: extracting and fusing the target dynamic interaction data in time and space to obtain time and space double-flow features; mining and expanding the unlabeled sample set through a positive unlabeled learning algorithm to obtain an expanded sample set of the unlabeled sample; training an initial core node prediction model based on the time and space double-flow features and the expanded sample set to obtain a target core node prediction model; and inferring and predicting the entity data to be predicted in the dynamic network through the target core node prediction model to obtain the core node in the dynamic network. The core node prediction model is trained by extracting and fusing the target dynamic interaction data in time and space and expanding the sample set in combination with the positive unlabeled learning algorithm, the problem that the prediction accuracy and the generalization ability are insufficient due to the time and space feature fragmentation and the positive sample scarcity is solved, and the core node in the dynamic network is more accurately predicted.

Description

Technical Field

[0001] This application relates to the field of data mining technology, and in particular to methods, apparatus, devices and storage media for predicting core nodes in dynamic networks. Background Technology

[0002] With the widespread application of complex dynamic network analysis in fields such as financial transactions, the Internet of Things in transportation, and large-scale social networks, the accurate identification of core nodes has become a crucial foundation for system risk assessment and optimal resource allocation. Existing technologies for node classification or prediction typically employ graph representation learning models to extract static topological features of the network, or sequence models to extract dynamic time-series features of node interactions, directly concatenating the two types of features before inputting them into a classifier for training. However, these existing technologies have a significant drawback: due to the fragmentation of spatiotemporal features and the fusion achieved through a simple concatenation method, the model fails to capture the nonlinear cross-relationships between network topology and dynamic interaction changes, resulting in a severely inadequate characterization of the true network influence of nodes.

[0003] The above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention

[0004] The main objective of this application is to provide a method, apparatus, device, and storage medium for predicting core nodes in dynamic networks, aiming to solve the technical problem of insufficient prediction accuracy for core nodes in dynamic networks.

[0005] To achieve the above objectives, this application proposes a method for predicting core nodes in a dynamic network, the method comprising: Spatiotemporal dual-stream extraction and fusion of the target dynamic interaction data are performed to obtain the spatiotemporal dual-stream features of the target dynamic interaction data; The unlabeled sample set is expanded by mining the unlabeled sample set through a positive example unlabeled learning algorithm to obtain the expanded sample set of the unlabeled samples; The initial core node prediction model is trained based on the spatiotemporal dual-stream features and the extended sample set to obtain the target core node prediction model. The core nodes in the dynamic network are obtained by reasoning and predicting the entity data to be predicted in the dynamic network through the target core node prediction model.

[0006] In one embodiment, the step of performing spatiotemporal dual-stream extraction and fusion on the target dynamic interaction data to obtain the spatiotemporal dual-stream features of the target dynamic interaction data further includes: Obtain the original dynamic interaction data of network entities, and perform data cleaning processing on the original dynamic interaction data; A multi-window sliding feature is constructed based on the cleaned original dynamic interaction data. The interaction weights of the multi-window sliding feature are then subjected to a logarithmic transformation to obtain the target dynamic interaction data.

[0007] In one embodiment, the step of constructing multi-window sliding features based on the cleaned original dynamic interaction data includes: Several time windows of different target lengths are set, and historical backtracking is performed based on the current time node to obtain the interaction frequency, total interaction weight and number of interacting neighbor nodes of network entities within the corresponding time window. The interaction frequency, the sum of interaction weights, and the number of interaction neighbor nodes obtained under the time window are combined to obtain the multi-window sliding feature corresponding to the time window.

[0008] In one embodiment, the step of performing spatiotemporal dual-stream extraction and fusion on the target dynamic interaction data to obtain the spatiotemporal dual-stream features of the target dynamic interaction data includes: Spatial flow features are extracted from the target dynamic interactive data to obtain spatial flow features; Temporal flow features are extracted from the target dynamic interactive data to obtain temporal flow features; The spatial flow features and the temporal flow features are concatenated to obtain the initial concatenated features; The initial spliced ​​features are input into an adaptive gated fusion network for cross-modal feature interaction and deep fusion to obtain the spatiotemporal dual-stream features.

[0009] In one embodiment, the step of extracting spatial flow features from the target dynamic interaction data to obtain spatial flow features includes: A directed graph is constructed using network entities as nodes and interaction relationships as directed edges, and a random walk sequence is generated on the directed graph; The random walk sequence is input into the target learning model for training, and the spatial flow features of the first preset dimension are output.

[0010] In one embodiment, the step of extracting time-flow features from the target dynamic interaction data to obtain time-flow features includes: Based on the interaction time, time interval, interaction type and interaction direction, the target dynamic interaction data is sequenced to obtain the entity time sequence. Linear features are obtained by linearly mapping the continuous numerical features in the time series of the entity; The discrete category features in the entity time sequence are represented by word embedding to obtain discrete features; The linear and discrete features are concatenated into a comprehensive feature vector, which is then input into a gated recurrent unit network based on time order. Extract the hidden state of the gated recurrent unit network at the last moment to obtain the time flow feature of the second preset dimension.

[0011] In one embodiment, the step of mining and expanding the unlabeled sample set using a positive example unlabeled learning algorithm to obtain an expanded sample set of the unlabeled samples includes: Reliable negative samples are selected from the unlabeled sample set; The known positive samples are labeled with the first value, and the reliable negative samples are labeled with the second value. The initial classifier is then trained. The remaining unlabeled samples are predicted by the trained initial classifier to obtain the predicted probability of the remaining unlabeled samples; Samples with predicted probabilities higher than a preset first threshold are added to the positive sample set as potential positive samples, and samples with predicted probabilities lower than a preset second threshold are added to the reliable negative sample set. Training and prediction are performed iteratively until the positive sample set no longer increases significantly or reaches the preset number of iterations, thus obtaining the expanded sample set.

[0012] Furthermore, to achieve the above objectives, this application also proposes a dynamic network core node prediction device, which includes: The feature extraction module is used to extract and fuse the target dynamic interaction data in both spatiotemporal streams to obtain the spatiotemporal dual-stream features of the target dynamic interaction data. The sample expansion module is used to mine and expand the unlabeled sample set through a positive example unlabeled learning algorithm to obtain an expanded sample set of the unlabeled samples; The model training module is used to train the initial core node prediction model based on the spatiotemporal dual-stream features and the extended sample set to obtain the target core node prediction model. The node prediction module is used to perform inference and prediction on the entity data to be predicted in the dynamic network through the target core node prediction model, so as to obtain the core nodes in the dynamic network.

[0013] In addition, to achieve the above objectives, this application also proposes a dynamic network core node prediction device, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the dynamic network core node prediction method as described above.

[0014] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the dynamic network core node prediction method described above.

[0015] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the dynamic network core node prediction method described above.

[0016] One or more technical solutions proposed in this application have at least the following technical effects: This application proposes a method, apparatus, device, and storage medium for predicting core nodes in dynamic networks. The method involves extracting and fusing spatiotemporal dual-stream data of a target dynamic interaction to obtain its spatiotemporal dual-stream features. An extended sample set of unlabeled samples is obtained by mining and expanding the unlabeled sample set using a positive example unlabeled learning algorithm. An initial core node prediction model is trained based on the spatiotemporal dual-stream features and the extended sample set to obtain a target core node prediction model. This model is then used to infer and predict entity data in the dynamic network, thereby obtaining the core nodes in the dynamic network. By extracting and fusing spatiotemporal dual-stream data of the target dynamic interaction and training the core node prediction model using a positive example unlabeled learning algorithm to expand the sample set, the method solves the problems of insufficient prediction accuracy and generalization ability caused by the fragmentation of spatiotemporal features and the scarcity of positive samples, thus more accurately predicting core nodes in dynamic networks. Attached Figure Description

[0017] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating an embodiment of the dynamic network core node prediction method of this application. Figure 2 This is a schematic diagram of the module structure of the dynamic network core node prediction device according to an embodiment of this application; Figure 3 This is a schematic diagram of the device structure of the hardware operating environment involved in the dynamic network core node prediction method in the embodiments of this application.

[0020] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0021] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0022] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0023] The main solution of this application embodiment is as follows: Spatiotemporal dual-stream extraction and fusion of target dynamic interaction data is performed to obtain the spatiotemporal dual-stream features of the target dynamic interaction data; the unlabeled sample set is mined and expanded using a positive example unlabeled learning algorithm to obtain an expanded sample set of the unlabeled samples; an initial core node prediction model is trained based on the spatiotemporal dual-stream features and the expanded sample set to obtain a target core node prediction model; and the target core node prediction model is used to infer and predict the entity data to be predicted in the dynamic network to obtain the core nodes in the dynamic network.

[0024] In this embodiment, for ease of description, the dynamic network core node prediction device will be used as the execution subject for the following description.

[0025] With the widespread application of complex dynamic network analysis in fields such as financial transactions, the Internet of Things in transportation, and large-scale social networks, the accurate identification of core nodes has become a crucial foundation for system risk assessment and optimal resource allocation. Existing technologies for node classification or prediction typically employ graph representation learning models to extract static topological features of the network, or sequence models to extract dynamic time-series features of node interactions, directly concatenating the two types of features before inputting them into a classifier for training. However, these existing technologies have a significant drawback: due to the fragmentation of spatiotemporal features and the fusion achieved through a simple concatenation method, the model fails to capture the nonlinear cross-relationships between network topology and dynamic interaction changes, resulting in a severely inadequate characterization of the true network influence of nodes.

[0026] This application provides a solution that trains a core node prediction model by extracting and fusing spatiotemporal dual-stream data of the target dynamic interaction data and expanding the sample set by combining positive example unlabeled learning algorithm. This solves the problems of insufficient prediction accuracy and generalization ability caused by spatiotemporal feature fragmentation and scarcity of positive samples, thereby more accurately predicting core nodes in dynamic networks.

[0027] It should be noted that the executing entity in this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an electronic device or dynamic network core node prediction device capable of performing the above functions. The following description uses a dynamic network core node prediction device as an example to illustrate this embodiment and the subsequent embodiments.

[0028] Based on this, embodiments of this application provide a method for predicting core nodes in a dynamic network, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the dynamic network core node prediction method of this application.

[0029] In this embodiment, the dynamic network core node prediction method includes steps S10-S13: Step S10: Perform spatiotemporal dual-stream extraction and fusion on the target dynamic interaction data to obtain the spatiotemporal dual-stream features of the target dynamic interaction data.

[0030] It should be noted that target dynamic interaction data refers to network entity interaction data used for feature engineering after preprocessing. It records the occurrence of each interaction with network entities as the main body, and its role is to provide a unified and regular data foundation for feature extraction in both spatial and temporal dimensions. Spatiotemporal dual-stream extraction and fusion refers to a processing method that extracts features from the same interaction data from both spatial and temporal dimensions, and then unifies the two sets of features. The spatial stream is used to characterize the structural position and connection relationship of network entities in the overall network topology, while the temporal stream is used to characterize the dynamic changes of the network entity's own interaction behavior over time. The technique used is dual-channel parallel feature extraction and feature fusion technology, that is, first decoupling heterogeneous information into two independent feature channels and modeling them separately, and then unifying them in the representation space. Spatiotemporal dual-stream features refer to the unified feature vector obtained after fusion, which carries both structural and temporal information. Its role is to serve as the sole input representation for subsequent sample mining and model training.

[0031] Understandably, the purpose of this step is to generate a unified representation of network entities that simultaneously encompasses topological structure and dynamic evolution. The influence of a network entity depends on both its connection position in the network (structural attribute) and the activity level and evolutionary trend of its interaction behavior (temporal attribute). If only one is considered, the representation will be incomplete and lack discriminative information. By extracting features in both spatial and temporal dimensions, each feature focuses on depicting the inherent laws of one attribute, avoiding the interference of two heterogeneous information streams in the same channel. Subsequently, the two features are fused into the same representation space, so that the fusion result retains discriminative information from both aspects. Its function is to make the obtained spatiotemporal dual-stream features more complete and sufficient in representational ability compared to single-dimensional features, thereby alleviating the problem of insufficient model representation of the true influence of nodes due to single representation dimension and fragmented spatiotemporal information. This provides a high-fidelity input foundation for subsequent sample mining and model training based on these features.

[0032] Specifically, firstly, a graph structure is constructed based on the interaction data, with entities as nodes and interaction relationships as edges. A graph representation learning approach is used to learn the structural vectors of each entity from the graph structure as spatial flow features. Simultaneously, the interaction records of each entity are organized into a sequence according to the interaction time order, and a sequence modeling approach is used to learn the temporal vectors of each entity from the sequence as temporal flow features. Then, the two vectors are concatenated along the feature dimension and fused. Spatial flow features can also be generated using a centrality index-based statistical method, and temporal flow features can also be generated using a time window-based statistical method, and then the two are fused. Fusion can be performed by directly outputting the concatenated features, or by applying learnable weights to the concatenated features for weighted fusion.

[0033] For example, taking a network entity as an example, its connection relationships across the entire network are statistically analyzed from its interaction data, generating a set of structural vectors (characterizing its neighbor size and location). Simultaneously, an interaction sequence is compiled according to the chronological order of its recent interactions, generating a set of temporal vectors (characterizing the changing trend of its interaction intensity). These two sets of vectors are then concatenated dimensionally into a longer vector. A set of weight coefficients, automatically learned by the network, is then applied to each dimension of the concatenated vector. This ensures that dimensions more important for determining whether the entity is a core node are retained, while less important dimensions are suppressed, ultimately outputting the entity's spatiotemporal dual-stream features. The principle behind applying the weight coefficients is as follows: the concatenated vector is input into a learnable mapping function, which outputs a set of values ​​with the same dimensions as the concatenated vector. These values ​​are then multiplied digit by digit of the concatenated vector, thereby achieving differentiated adjustment of the importance of each dimension without changing the dimensions themselves. This transforms the fusion process from fixed-rule concatenation to data-driven adaptive weighting.

[0034] Step S11: The unlabeled sample set is expanded by mining using a positive example unlabeled learning algorithm to obtain an expanded sample set of the unlabeled samples.

[0035] It should be noted that the positive-unlabeled learning algorithm is a semi-supervised learning algorithm that learns under the condition that only a small number of positive samples are known and the categories of the remaining samples are unknown. The technique it uses is positive-unlabeled learning (PU learning), that is, it does not treat all unlabeled samples as negative classes, but infers the sample category by using the relationship between unlabeled samples and known positive samples. The unlabeled sample set is a collection of network entity samples whose category labels are unknown, and its role is to provide a source of candidate samples for mining. Mining expansion is the process of discovering and confirming potential similar samples in the unlabeled samples and adding them to the labeled set. The expanded sample set is a sample set obtained after mining, which has expanded the number of samples and category coverage compared to the initial positive sample set. Its role is to provide training samples with sufficient scale and more complete labels for subsequent model training.

[0036] Understandably, the purpose of this step is to construct a usable training sample set for model training in scenarios where positive samples are extremely scarce. If a massive number of unlabeled samples are directly treated as negative classes for training, the actual positive samples will be incorrectly treated as negative classes, introducing a large amount of noise labeling, causing the decision boundary learned by the classifier to shift and fall into overfitting. The positive example unlabeled learning algorithm only uses known positive classes as confirmed positive examples, and does not make an overall negative assumption for unlabeled samples. Instead, it infers based on the distribution and similarity relationships between unlabeled samples and known positive samples, gradually confirming the samples with higher confidence as a certain category and including them in the label set. Its function is to effectively expand the scale and category coverage of the training samples, while avoiding the destruction of the classification boundary by erroneous negative class labeling, thereby alleviating the problems of poor model generalization ability and decision boundary shift caused by the scarcity of positive samples.

[0037] Specifically, first, samples that are certain not to belong to the positive class are selected from the unlabeled sample set based on their difference from known positive class samples, and these are designated as negative classes. These negative classes, along with the known positive classes, constitute the initial labeled data. A base classifier is then trained using this classifier. This classifier is then used to score the remaining unlabeled samples and incorporate them into the labeled set in batches according to their confidence level, iteratively updating the classifier. Alternatively, a clustering-based approach can be directly adopted, stratifying the unlabeled samples in the feature space based on their distance from known positive class samples, and designating the samples furthest from the positive cluster as negative classes. When iteratively incorporating samples, a one-sided threshold can be used for gradual inclusion, or a two-sided threshold can be used to simultaneously include both positive and negative classes.

[0038] For example, suppose a network has several known core nodes (positive class) and a large number of nodes of unknown class. First, calculate the distance between each unknown node and the known core nodes in the feature space. Identify the group of unknown nodes with the farthest distance as reliable non-core nodes (negative class). Train a base classifier using the known core nodes as the positive class and this group of nodes as the negative class. Then, use this classifier to score the remaining unknown nodes one by one to obtain their probability values ​​of belonging to core nodes. Add nodes with high probabilities to the positive class and nodes with low probabilities to the negative class. Retrain the classifier based on the expanded positive and negative classes and repeat the above scoring and addition process until the number of newly added positive nodes is very small or a predetermined number of rounds is reached. The resulting set of positive and negative classes is the expanded sample set. The scoring calculation principle is as follows: the classifier takes the aforementioned spatiotemporal dual-stream features as input and outputs a posterior probability estimate of the sample belonging to the positive class. The higher the probability, the closer the sample is to the known positive class distribution in the feature space, and therefore the higher the confidence in including it in the positive class.

[0039] Step S12: Train the initial core node prediction model based on the spatiotemporal dual-stream features and the extended sample set to obtain the target core node prediction model.

[0040] It should be noted that the initial core node prediction model refers to the node classification model that has not yet completed training and whose parameters are in the initial state. Its role is to serve as a baseline model to be optimized, and to gradually adapt its parameters to the sample distribution of this scheme through training. The expanded sample set refers to the set of labeled samples obtained after mining and expansion in step S11. Its role is to provide training with a sufficiently large and relatively complete set of positive and negative class supervision signals. Training refers to the parameter optimization process that takes sample features as input and sample labels as the target, measures the deviation between the prediction results and the labels through a loss function, and adjusts the model parameters in reverse to gradually reduce the deviation. The technique used is supervised parameter learning based on loss function and backpropagation. The target core node prediction model refers to the model whose parameters have converged after training and can be directly used for inference and prediction. Its role is to carry the classification ability obtained from training and to make judgments on unknown samples.

[0041] Understandably, the purpose of this step is to transform the results of the previous two steps—high-fidelity spatiotemporal dual-stream features and an expanded sample set—into a predictive model with practical discriminative capabilities. The model's discriminative ability depends on the quality of the input feature representation and the quality of the training sample annotations. The former determines whether the classification boundary is separable, and the latter determines whether the classification boundary is correctly characterized. Using spatiotemporal dual-stream features that simultaneously carry structural and temporal information as input, the samples have better separability in the feature space. Using a mined, expanded, and more reliably labeled sample set as a supervision signal, the model can learn sufficiently on both positive and negative samples. By measuring the prediction bias through a loss function and iteratively optimizing the parameters, the model gradually approaches the true classification boundary. Its function is to enable the trained target model to simultaneously possess the ability to utilize structural and temporal information and robust discriminative ability under conditions of a small number of positive samples, thereby alleviating the prediction inaccuracy problem caused by insufficient feature representation and sample scarcity.

[0042] Specifically, the spatiotemporal dual-stream features corresponding to each sample in the expanded sample set are organized into a feature matrix, with sample labels as tags. A gradient boosting decision tree model is used for training, and the overall loss is optimized by fitting the residuals of each decision tree. In one embodiment of this application, a deep neural network is used to construct the model, and the parameters are optimized through multi-layer nonlinear transformations and backpropagation. During training, a cross-entropy loss function can be used to measure classification bias, or a weighted loss function that takes into account class imbalance can be used.

[0043] For example, the system extracts the spatiotemporal dual-stream feature vector corresponding to each sample in the extended sample set, and pairs it with the label indicating whether it is a core node (positive class is denoted as 1, negative class as 0) to form training sample pairs. During training, the model takes the feature vector as input and outputs a probability value belonging to the positive class. This probability value and the labeled value are substituted into the loss function to calculate the deviation (the deviation increases as the probability value deviates from the labeled value). Then, based on this deviation, the internal parameters of the model are adjusted along the gradient direction so that the probability value output in the next round is closer to the labeled value. This process is repeated for all training samples and several rounds until the loss value no longer decreases significantly. At this point, the model parameters are saved, thus obtaining the target core node prediction model. The calculation principle of parameter adjustment is as follows: the partial derivative of the loss function with respect to the model parameters yields the gradient. The gradient indicates the direction of parameter change that causes the loss to increase the fastest. Therefore, updating the parameters in the opposite direction of the gradient with a certain step size can gradually reduce the loss. This iterative process achieves a gradual approximation of the true classification boundary.

[0044] Step S13: Use the target core node prediction model to perform inference prediction on the entity data to be predicted in the dynamic network to obtain the core nodes in the dynamic network.

[0045] It should be noted that the data of entities to be predicted refers to the interaction data of network entities in a dynamic network that have not yet been determined to be core nodes. Its role is to serve as the input object in the inference stage. Inference prediction refers to the process of inputting the data of entities to be predicted into the trained model after undergoing the same processing as in the training stage, and having the model output the discrimination result. The technology used is forward inference technology based on the trained model. Core nodes refer to nodes in a dynamic network that have strong radiation and resource aggregation capabilities and play a key role in network connectivity and resource flow. Their role is to serve as the target result to be output by this method, for use in downstream risk assessment, resource allocation, and anomaly monitoring.

[0046] Understandably, the purpose of this step is to apply the trained model to the actual network entities to be judged, and output the final core node identification result. During the training phase, the model has learned the discrimination rules for distinguishing core nodes from non-core nodes from the expanded sample set, and these rules are solidified in the form of model parameters. During the inference phase, as long as the data of the entity to be predicted is subjected to the same processing as the training samples (i.e., the same spatiotemporal dual-stream extraction and fusion), and the feature expression is isomorphic to the training samples is obtained, the model can transfer the discrimination rules learned in the training phase to the entity and output the probability that it belongs to the core node. The category can be determined by comparing this probability with the judgment criteria. Its role is to make the whole method form a closed loop of feature extraction - sample expansion - model training - inference judgment, so that the training capability is truly put into practice as the ability to judge unknown entities, thereby stably identifying core nodes with connectivity control in dynamic networks.

[0047] Specifically, the data of entities to be predicted undergoes the same feature extraction process as the training samples to obtain their spatiotemporal dual-stream features. These features are then input into the target model to obtain probability values. The probability values ​​are compared with a decision threshold to output the decision result. In online scenarios, features can be extracted for a batch of entities to be predicted in one go, and then input into the model for batch inference. In offline scenarios, the entire process can be periodically re-executed to update the model and decision results.

[0048] For example, for a network entity whose status as a key node has not yet been determined, the system first performs spatiotemporal dual-stream extraction and fusion on its interaction data in the same manner as the training samples to obtain the spatiotemporal dual-stream feature vector of the entity. This vector is then input into the target core node prediction model for forward computation. The model performs linear and nonlinear transformations along each layer's parameters, ultimately outputting the probability value that the entity belongs to a core node. The system compares this probability value with a predetermined threshold. If the probability value is higher than the threshold, the entity is determined to be a core node and output; otherwise, it is determined to be a non-core node. The principle of forward computation is as follows: the input features and the weights of each layer of the model are sequentially multiplied by a matrix and biased, then transformed by a nonlinear function and passed layer by layer. Finally, at the output layer, it is mapped to a probability with a value between 0 and 1, which reflects the degree to which the sample is close to the positive class region in the feature space.

[0049] This embodiment, through the aforementioned scheme, first constructs a unified representation at the network entity level that simultaneously carries structural and temporal information, enabling the previously fragmented spatial and temporal information to complement each other in the same representation space, thereby improving the separability and expressive power of features. Then, at the sample level, it infers and expands labels from a massive number of unlabeled samples using positive examples without labels, avoiding the introduction of noise and disruption of classification boundaries by treating unknown samples as negative classes. This ensures that even with scarce positive samples, it can still provide training with a sufficient number of reliable labeled samples. Subsequently, training and inference share the same feature processing flow, ensuring that the discrimination rules obtained during training can be transferred losslessly to unknown entities. Thus, it improves the model's discrimination foundation from the two root causes of representation quality and sample quality, alleviating the problems of poor generalization and decision boundary offset caused by the fragmentation of spatiotemporal features and the scarcity of positive samples, making the output nodes closer to real core nodes in terms of convergence and connectivity control.

[0050] Based on the above implementation scheme, in one feasible implementation, the step of extracting and fusing spatiotemporal dual-stream data of the target dynamic interaction data to obtain the spatiotemporal dual-stream features of the target dynamic interaction data further includes S20~S21 before the step of extracting and fusing spatiotemporal dual-stream data of the target dynamic interaction data: Step S20: Obtain the original dynamic interaction data of the network entities and perform data cleaning processing on the original dynamic interaction data.

[0051] It should be noted that network entities refer to the main units that interact in a dynamic network, and are the basic objects for subsequent feature extraction and node determination; raw dynamic interaction data refers to the collection of interaction records that are directly collected and unprocessed, which usually contains fields such as node identifier, interaction time, interaction direction, interaction weight, and interaction type, and serves as the data source for the entire method; data cleaning refers to the process of identifying and removing defective records in the raw data to make the data regular and usable. The technology used is the abnormal record removal and isolated object filtering technology in data quality governance, that is, setting criteria based on field integrity and entity connectivity, and excluding records or entities that do not meet the criteria from subsequent calculations.

[0052] Understandably, the purpose of this step is to provide a reliable and well-organized data foundation for all subsequent feature calculations and model training. The original data contains abnormal records with missing key fields and isolated entities that do not interact with any other entities. The former will distort the statistics and destroy the integrity of feature calculations, while the latter has neither structural information nor temporal evolution information, and including them in the calculation will only introduce noise and dilute the effective samples. By removing these two types of objects before feature extraction, each entity participating in the subsequent calculation can have complete fields and real interaction relationships, so that the statistics reflect the real interaction situation and the graph structure and temporal sequence are built on effective entities. Its function is to improve the quality and consistency of input data, avoid dirty data from being propagated along the process and amplified into feature bias and model bias, thereby ensuring the reliability of subsequent spatiotemporal dual-stream features and sample expansion results.

[0053] Specifically, for each interaction record, check whether its key fields are complete, and remove records that are missing any key field such as node identifier, interaction time, or interaction weight. Count the number of times each network entity appears in the interaction records in the full dataset, and remove entities that have never appeared in any interaction record, or whose interaction records have been removed and are no longer connected to any entity, as isolated entities. Records with invalid interaction weights (such as null or negative values) can also be removed.

[0054] Step S21: Construct multi-window sliding features based on the cleaned original dynamic interaction data, and perform logarithmic transformation on the interaction weights of the multi-window sliding features to obtain the target dynamic interaction data.

[0055] It should be noted that the cleaned original dynamic interaction data refers to the interaction data retained after removing abnormal records and isolated entities. Its purpose is to serve as the input for the statistics and transformation in this step. Multi-window sliding features refer to the feature set obtained by setting statistical windows with different time spans and statistically analyzing the interaction of entities within each window. The technique used is multi-timescale sliding window statistical technology, which involves backtracking to multiple intervals of different lengths from the current time as a baseline and aggregating them separately. Its purpose is to simultaneously characterize the activity level and evolution trend of entities in the near, medium and long term. Interaction weights refer to the quantitative indicators (such as amount, flow, etc.) corresponding to each interaction, reflecting the scale or intensity of a single interaction. Logarithmic transformation processing refers to the numerical transformation of the interaction weight values ​​by applying a logarithmic function to compress the scale. The technique used is long-tail distribution compression technology based on logarithmic function. Target dynamic interaction data refers to the data obtained after completing multi-window statistics and logarithmic transformation, which can be directly entered into the feature extraction stage.

[0056] Understandably, the purpose of this step is to organize the scattered raw interaction records into standardized data that encompasses active information across multiple time scales and has a numerical distribution suitable for model learning. A single interaction record can only reflect instantaneous behavior, while whether an entity is a key node often depends on its continuous performance over a period of time. By backtracking and aggregating with time windows of different lengths, we can simultaneously obtain short-term current activity, medium-term development trends, and long-term stable scale. The multiple windows complement each other, avoiding the limitation that single-section data can only reflect the state at a certain moment. On the other hand, the interaction weights in the network usually follow a long-tail distribution, meaning that the values ​​of a few interactions are much larger than those of most interactions. If the raw values ​​are directly used in subsequent calculations, these maxima will dominate distance metrics and gradient updates, masking the distribution patterns of most samples. Logarithmic transformation can compress the value span while maintaining the order of magnitude. Its function is to make the output target dynamic interaction data have multi-scale expressiveness in the time dimension and a distribution characteristic suitable for learning in the numerical dimension, thereby providing higher quality input for the subsequent extraction of spatiotemporal dual-stream features.

[0057] Specifically, several time windows of varying lengths are first defined. Using the current time point as a baseline, data is backdated historically. Within each window, several interaction statistics for the entities are collected. These statistics from each window are then combined into a multi-window sliding feature. Subsequently, a logarithmic transformation is applied to the interaction weights within the feature. Furthermore, the logarithmic transformation can be a logarithmic form with a base of 1 offset, or a logarithmic form with other bases. In selecting the multi-window statistics, both statistics reflecting frequency and statistics reflecting scale can be used simultaneously.

[0058] This embodiment employs the above-described scheme, first by cleaning and removing abnormal records and isolated entities with missing fields, ensuring that every entity participating in subsequent calculations possesses complete fields and genuine interaction relationships, thus preventing dirty data from being amplified along the process. Next, it uses multiple time windows of varying lengths for backtracking and aggregation, enabling features to simultaneously carry active information at short-term, medium-term, and long-term scales, overcoming the limitation of single-section data only reflecting instantaneous states. Finally, it uses logarithmic transformation to compress the long-tail span of interaction weights, preventing maxima from dominating subsequent distance metrics and parameter updates without altering the order of magnitude. Therefore, it provides spatiotemporal dual-stream extraction and fusion with input data that is complete in fields, valid in entities, multi-level in scale, and with a regular numerical distribution, allowing subsequent extracted structural and temporal features to be built on a more reliable data foundation, indirectly supporting the resolution of the problems of spatiotemporal feature fragmentation and the scarcity of positive samples.

[0059] Based on the above implementation scheme, in one feasible implementation, the step of constructing multi-window sliding features based on the cleaned original dynamic interaction data includes S30~S31: Step S30: Set several time windows of different target lengths, and perform historical backtracking based on the current time node and the time windows to obtain the interaction frequency, total interaction weight and number of interaction neighbor nodes of network entities within the corresponding time windows.

[0060] It should be noted that a time window refers to a fixed-length time interval intercepted from a certain point in time. Different target lengths refer to the different time spans of multiple windows, which are set to cover the short-term, medium-term, and long-term levels respectively. The current time node refers to the baseline time used when performing statistics, which provides a unified backtracking starting point for each window. Historical backtracking refers to the operation of intercepting an interval along the time axis from the current time node and retrieving the data within that interval, using the sliding window interval intercepting technique. Interaction frequency refers to the number of times the entity participates in interactions within the window. The total interaction weight refers to the sum of all interaction weights of the entity within the window. The number of interaction neighbor nodes refers to the number of different peer entities that have interacted with the entity within the window. The three factors are used to characterize the activity status of the entity from three aspects: frequency, scale, and connection breadth.

[0061] Understandably, the purpose of this step is to measure the interaction status of entities across multiple time intervals with different spans, thereby obtaining basic statistics reflecting the activity characteristics at different time scales. An entity's current influence is not consistent with its performance in the near, medium, and long term—near-term data reflects current activity and sudden changes, while longer-term data reflects stable scale and sustainability. If only a single window is used, it can only depict one level and cannot distinguish between different situations such as recent activity but long-term stagnation and long-term stability but recent decline. By setting several windows of different lengths and backtracking from the same current moment, multiple scales of measurement results can be obtained simultaneously in one statistical analysis, enabling subsequent features to distinguish between the above-mentioned different situations. Its function is to make the obtained statistics hierarchical and complementary in the time dimension, providing a directly usable numerical basis for subsequently combining multi-window sliding features with complete information.

[0062] Specifically, three time windows of increasing length are set, corresponding to short-term, medium-term, and long-term, respectively. For example, the short-term window is set to 3 months, the medium-term window to 6 months, and the long-term window to 12 months. The aforementioned three statistics are calculated for each network entity within each time window. The time windows can be divided using a fixed span of calendar months or calendar days. When counting the number of interacting neighbor nodes, the count can be performed by deduplicating the peer entity identifier, or by counting separately according to the direction of interaction and then merging the results.

[0063] Step S31: Combine the interaction frequency, total interaction weight, and number of interaction neighbor nodes obtained under the time window to obtain the multi-window sliding feature corresponding to the time window.

[0064] It should be noted that combination refers to the operation of arranging and merging values ​​scattered in different windows and different statistical aspects into a whole in a predetermined order. The technique used is the feature vector dimensional concatenation technique, that is, arranging each value as a dimension of the vector in sequence. Multi-window sliding features refer to the multi-scale active features obtained after combination and expressed in vector form. Their role is to serve as input for logarithmic transformation and further constitute the target dynamic interactive data, which is ultimately used for spatiotemporal dual-stream feature extraction.

[0065] Understandably, the purpose of this step is to integrate scattered statistics from multiple windows and perspectives into a well-structured feature representation that can be directly used in subsequent steps. Since each statistic is independent in terms of numerical dimensions and meaning, their scattered arrangement prevents them from directly participating in subsequent distance calculations and model training. Only by organizing them into a unified vector object in a defined order can an entity correspond to a feature, providing a consistent interface for subsequent logarithmic transformations, feature extraction, and model input. Simultaneously, placing the values ​​from different windows and perspectives side-by-side on different dimensions of the same vector allows subsequent processing to simultaneously view all information about the same entity across multiple scales and perspectives, rather than using them in isolation. Its function is to transform multi-scale, multi-faceted statistical results into a unified, computable feature representation, thereby making the characterization of entity activity both complete and well-structured.

[0066] Specifically, in one embodiment of this application, the values ​​are arranged sequentially as dimensions of a vector according to the window-statistic order, that is, the three statistics under one window are arranged first, then the three statistics under the next window are arranged, until all windows are arranged. In another embodiment of this application, the statistics are arranged in the order of statistic-window, that is, the interaction frequency under all windows is arranged first, then the sum of interaction weights under all windows is arranged, and finally the number of interactive neighbor nodes under all windows is arranged. When combining, the statistics can be arranged as is, or the statistics can be normalized first and then arranged.

[0067] This embodiment, through the aforementioned scheme, first sets several time windows of different lengths and backtracks from the current moment as a baseline, acquiring the interaction frequency, total interaction weight, and number of interacting neighbor nodes within each window. Then, it combines the statistics from each window into a multi-window sliding feature. This allows the activity status of the same entity to be measured simultaneously from three aspects: frequency, scale, and connection breadth, and across three scales: short-term, medium-term, and long-term. This overcomes the limitation of single-section data, which can only reflect instantaneous states and cannot distinguish different evolutionary scenarios. Furthermore, by combining scattered statistics sequentially, it integrates them into a vector form with fixed positional meaning and unified computation. This provides a clear and feasible construction method for multi-window sliding features, enabling the target dynamic interaction data to possess a hierarchical, complementary, and structurally regular activity expression in the time dimension. It provides a reliable statistical basis for subsequent spatiotemporal dual-stream feature extraction and sample selection in positive example unlabeled learning, thereby supporting the solution to the problems of spatiotemporal feature fragmentation and positive sample scarcity.

[0068] Based on the above implementation scheme, in a feasible implementation, the step of extracting and fusing spatiotemporal dual-stream data of the target dynamic interaction data to obtain the spatiotemporal dual-stream features of the target dynamic interaction data includes S40~S43: Step S40: Extract spatial flow features from the target dynamic interactive data to obtain spatial flow features.

[0069] It should be noted that spatial flow feature extraction is a process of learning entity vector representations from the connection relationships between network entities (i.e., the graph structure of the network). The technique used is graph representation learning, which describes the network as a graph composed of nodes and edges and learns the low-dimensional dense vector of each node. Spatial flow features are feature vectors obtained from this process, which describe the position and connection status of entities in the network topology. Their role is to characterize the structural attributes of entities and serve as an input for subsequent splicing and fusion.

[0070] Understandably, the purpose of this step is to generate a vector representation of an entity's topological position and connectivity at the network structural level. The influence of a network entity is closely related to its connectivity within the network—entities in pivotal positions with extensive connections often possess stronger radiation and convergence capabilities. By organizing all entities and their interactions into a graph structure and learning vector representations from it, high-dimensional, sparse, and difficult-to-computate connectivity relationships can be mapped into low-dimensional, dense vectors that can be directly used in computation, allowing structural attributes to be quantified. Its function is to transform the structural attributes of entities into feature forms that can be subsequently spliced, fused, and used by the model, thereby enabling the final fused features to possess discriminative information at the structural level.

[0071] Specifically, a graph structure is first constructed with entities as nodes and interactions as edges. Then, node sequences are obtained by sampling on the graph, and node vectors are learned using these sequences as training corpus. Alternatively, node vectors can be obtained by directly aggregating neighbor information layer by layer on the graph without sequence sampling. In graph construction, interactions can be treated as directed edges to retain directional information, or as undirected edges to simplify the structure.

[0072] Step S41: Extract time-flow features from the target dynamic interaction data to obtain time-flow features.

[0073] It should be noted that temporal flow feature extraction refers to the process of learning the feature representation of a network entity from the sequence of its own interactive behavior over time. The technique used is sequence modeling, which treats the interaction records of entities arranged in time as a sequence and passes the state step by step along the sequence. Temporal flow features refer to the feature vectors obtained from this process that characterize the dynamic evolution of entity interactive behavior. Their role is to represent the temporal attributes of the entity and serve as another input for subsequent splicing and fusion.

[0074] Understandably, the purpose of this step is to generate a vector representation of the dynamic changes in the interactive behavior of an entity from a temporal perspective. Whether an entity is a key node depends not only on its current state but also on the changing trend of its interactive behavior over time—even if a continuously active entity and a gradually decaying entity are of similar current size, their subsequent influence may be drastically different. By organizing the entity's interactions into a sequence according to time and accumulating states step by step along the sequence, the final representation can carry information about how past behaviors affect the current state, thereby encoding the trend of change into a vector. Its function is to transform the temporal attributes of the entity into a feature form that can be used by subsequent splicing, fusion, and modeling, thus enabling the final fused features to have temporal discriminative information.

[0075] Specifically, the interactions of an entity are arranged chronologically into a sequence, with each position in the sequence carrying relevant fields for that interaction. This sequence is then progressively fed into a recurrent network, and the state at the end of the sequence is taken as the entity's vector. Alternatively, a recurrent structure can be avoided, and vectors can be obtained through convolution or attention convergence along the time dimension. The sequence elements can include numerical information such as the scale and interval of the interaction, or categorical information such as the type and direction of the interaction.

[0076] Step S42: Concatenate the spatial flow feature with the temporal flow feature to obtain the initial concatenated feature.

[0077] It should be noted that spatial flow features and temporal flow features refer to two vectors that characterize structural attributes and temporal attributes, respectively. Their role is to provide discriminative information at the structural and temporal levels, respectively. Concatenation refers to the operation of joining two vectors end to end along a dimension to form a longer vector, using the technique of vector concatenation along dimensions. The initial concatenation feature refers to the combined vector obtained after concatenation but before fusion processing, which serves as the input for deep fusion, providing a unified and dimensionally defined input object for cross-modal interaction.

[0078] Understandably, the purpose of this step is to merge the two heterogeneous features from the structural and temporal channels into a unified object, providing input for subsequent deep fusion. Structural and temporal features are generated independently and belong to different representation spaces. Only by placing them on different dimensional segments of the same vector can we perform unified calculations and adjustments on which dimensions should be emphasized and which should be suppressed. The splicing operation itself does not change the original values ​​of each dimension, but rather places the two features in a defined order, so that the fusion stage can see both information simultaneously within the same vector and learn their interrelationships. Its function is to construct a unified input with defined dimensions, clear order, and direct access to the fusion stage, thus providing the premise for joint processing of structural and temporal information.

[0079] Specifically, in one embodiment of this application, spatial flow features are placed at the beginning of the vector and temporal flow features are placed at the end of the vector, and then directly concatenated in this order. In another embodiment of this application, temporal flow features are placed at the beginning and spatial flow features are placed at the end. Before concatenation, the two feature streams can be length aligned or scale normalized respectively, and then concatenated.

[0080] Step S43: Input the initial spliced ​​features into an adaptive gated fusion network to perform cross-modal feature interaction and deep fusion to obtain the spatiotemporal dual-stream features.

[0081] It should be noted that adaptive gating fusion network refers to a learnable network module that can automatically generate weights for each dimension based on the content of the input sample and adjust the input accordingly. The technology used is soft gating weighting based on neural networks, that is, the network first outputs a set of weights with the same dimension as the input, and then adjusts the input dimension by dimension with these weights. Cross-modal feature interaction refers to the mutual calculation and influence between dimensions that originally belong to the structural channel and the temporal channel, rather than each participating in subsequent stages independently. Deep fusion refers to further integrating information on the basis of interaction and outputting a unified fusion result. Spatiotemporal dual-stream features refer to the final features obtained after fusion, which serve as sample representations in positive unlabeled learning and as inputs for model training and inference.

[0082] Understandably, the purpose of this step is to further process the concatenated vector, enabling the information from the structural and temporal channels to truly interact and adaptively integrate according to the sample. Simple concatenation assumes that each dimension has equal importance and is independent of the others; this assumption does not hold true in complex networks—the structural and temporal information of the same entity contribute differently to the decision in different contexts, and there is an interaction between the two (for example, an entity in a pivotal position in the topology but whose interaction has recently diminished needs to have its true influence remeasured). By feeding the concatenated vector into a learnable gated network, the network can then perform further processing based on the sample. The concatenated vector itself generates a set of weights corresponding to each dimension. These weights are then used to adjust the concatenated vector dimension by dimension. This allows the generation of weights to depend on the sample content rather than fixed rules, thus enabling different samples to receive different weighting methods. At the same time, since the weight generation process performs unified calculations on all dimensions, the structural and temporal dimensions can influence each other in the same calculation, achieving cross-channel interaction. Its function is to transform the fusion result from mechanical superposition to sample-adaptive cross-modal integration, which retains the truly discriminative dimensions, suppresses redundant dimensions, and captures the nonlinear cross-relationships between the two channels, thereby improving the expressive power of the fusion features.

[0083] Specifically, the concatenated vector is input into a multilayer perceptron for cross-dimensional computation, then processed by an activation function that compresses the values ​​to between 0 and 1, resulting in weights with the same dimensions as the concatenated vector. Finally, these weights are multiplied dimension-by-dimensionally by the concatenated vector to obtain the fusion result. For the activation function, an sigmoid function that compresses values ​​to the 0-1 range can be used, or a normalization function that treats values ​​as relative proportions can be used. In terms of network structure, a two-layer perceptron can be used to achieve stronger fitting ability, or a single-layer mapping can be used to reduce complexity.

[0084] This embodiment employs the aforementioned scheme, first extracting spatial and temporal features from the target dynamic interaction data separately, then concatenating the two feature streams into an initial concatenated feature, and finally feeding it into an adaptive gated fusion network for cross-modal feature interaction and deep fusion. This ensures that structural and temporal information are fully modeled in their respective adapted channels, avoiding interference between heterogeneous information within the same channel. They are then juxtaposed along the same vector dimension, and subsequently, in the gated network, dimension-wise weights are generated driven by the samples themselves, completing cross-channel interaction and integration. Thus, the spatiotemporal dual-stream extraction and fusion is implemented as an executable link of dual-path decoupling—dimensional juxtaposition—gated weighting. This transforms fusion from a mechanical superposition of dimensions with equal importance and independence into an adaptive process that highlights discriminative dimensions, suppresses redundant dimensions, and captures the nonlinear cross-relationship between structure and temporal sequence based on sample content. This significantly enhances the ability of fused features to characterize the real network influence of nodes, directly addressing the technical problems of fragmented spatiotemporal features and the inability of simple concatenation to capture cross-modal cross-information.

[0085] Based on the above implementation scheme, in one feasible implementation, the step of extracting spatial flow features from the target dynamic interaction data to obtain spatial flow features includes S50~S51: Step S50: Construct a directed graph with network entities as nodes and interaction relationships as directed edges, and generate a random walk sequence on the directed graph.

[0086] It should be noted that network entities are the basic objects that constitute the network, and in this step, they are treated as nodes in the graph; interaction relationships refer to the fact that two entities have interacted, and in this step, they are treated as edges in the graph; directed edges are edges with directional attributes, with their direction pointing from the initiator of the interaction to the receiver. Their function is to preserve the active and passive semantics of the interaction, enabling the graph to distinguish the different roles of entities in the interaction; a directed graph is a graph structure composed of the aforementioned nodes and directed edges. Its function is to provide traversable structural objects for subsequent walk sampling, and it is a key intermediate product for transforming the original interaction data into a computable topological structure; a random walk sequence is a sequence of nodes traversed by moving randomly hop by hop along the edges on the graph. The technique used is random walk sampling on the graph, and its function is to transform the topological proximity relationships in the graph, which are difficult to model directly, into a linear sequence form that can be processed by the sequence model.

[0087] Understandably, the purpose of this step is to transform the implicit topological structure of the original interaction data into a sample form that can be directly trained by subsequent models. Graph structures are non-linear and non-sequential data forms, making them difficult to directly input into conventional representation learning models. Random walks, by moving hops along the edges of the graph, transform the proximity and structural similarity of nodes in the graph into co-occurrence relationships in the sequence—nodes that are close to each other or structurally similar in the graph will appear in the same walk sequence with a higher probability. Thus, the problem of learning the structure on the graph is transformed into the problem of learning co-occurrence in the sequence, which can be efficiently handled by mature representation learning models. At the same time, directed edges are used instead of undirected edges, so that the walk process follows the actual direction of the interaction, avoiding the indiscriminate treatment of the initiator and the receiver. Its role is to construct a corpus that retains both directional and hierarchical structural information and can be directly trained by sequence models, laying the foundation for the next step of learning spatial flow features.

[0088] Specifically, starting from each network entity, several walks are performed. Each walk moves hop-by-hop along directed edges to a preset number of steps, and the nodes visited are recorded as a sequence. The sequences of all starting points and all walks are compiled into the training corpus. Regarding the direction of the walk, movement can be strictly along the direction of directed edges, or movement along edges in an undirected sense can be allowed to expand the coverage. For the walk strategy, a uniformly random selection of the next hop from the outgoing edges of the current node can be used, or a biased approach can be adopted (balancing different inclinations between returning to the coming path and exploring outwards).

[0089] For example, first, assign a node to each entity and register each interaction as a directed edge from initiator to receiver, forming a directed graph. Then, starting from node D, examine all its outgoing edges and randomly select one, move to the next node pointed to by that edge, and repeat the process of examining outgoing edges, randomly selecting, and moving at the new node. Perform this process several times, recording the node identifiers along the way, thus obtaining a random walk sequence starting with node D. Repeat this process several times for node D and traverse all nodes, summing up the entire random walk sequence. The calculation principle of the walk is that the movement of each hop depends only on the set of outgoing edges of the current node and the selection probability on that set, and is independent of nodes visited earlier. Therefore, this process constitutes a progressively advancing chain of random selections. Since the probability of repeatedly traversing the same edge is related to the connection strength and co-neighbor structure of the two endpoints in the graph, the co-occurrence frequency of nodes in the final sequence implicitly contains the proximity and structural similarity relationships of the original graph, thus providing a basis for learning node vectors using the sequence as corpus.

[0090] Step S51: Input the random walk sequence into the target learning model for training and output the spatial flow features of the first preset dimension.

[0091] It should be noted that the target learning model refers to a graph representation learning model used to learn node vector representations from a sequence. The technique it employs is a representation learning technique based on sequence co-occurrence (specifically including the jump word model that predicts context nodes with the center node and its improved forms, which learns vectors by constructing center-context pairs on the sequence and optimizing their similarity). Its role is to transform the obtained sequence corpus into a dense vector for each node. Training refers to the parameter learning process of repeatedly adjusting node vectors by optimizing the target using paired samples constructed from the sequence as the object, so as to gradually improve the objective function. The first preset dimension refers to the pre-set dimension length of the spatial flow feature vector. Spatial flow features refer to the fixed-length vector corresponding to each node output after training, which is used to characterize the topological position and centrality of the nodes.

[0092] Understandably, the purpose of this step is to solidify the implicit node co-occurrence relationships in the random walk sequence into directly computable low-dimensional vectors. Nodes that appear close to each other in the sequence often have proximity or structural similarity relationships in the graph. If this proximity co-occurrence can be reflected as close distances in the vector space, then the node vectors encode the topological structure of the graph. The model constructs several central node-context node pairings through a sliding window on the sequence, and uses the nodes within the pair as positive examples and randomly selected nodes as negative examples. The optimization increases the vector similarity of positive example pairings and decreases the similarity of negative example pairings, thereby compressing the co-occurrence statistics into the vector of each node. Its function is to output spatial flow features with uniform dimension, comparable density, and carrying topological and centrality information, so that they can be directly concatenated with temporal flow features in dimension and enter into subsequent fusion and model training.

[0093] Specifically, using the jump-word model as a framework, a sliding window is used to construct training samples by taking the center node and its context nodes in each sequence. Negative examples are constructed using negative sampling, and node vectors are updated through gradient optimization. Regarding the walking strategy, it can be combined with the aforementioned biased walking, making the learned vectors adjustable between local proximity similarity and structural role similarity. In terms of dimension setting, the first preset dimension can be set to... For example, all random walk sequences are aggregated into a corpus, and a fixed-length window is slid across each sequence: the node corresponding to the center position of the window is the center node, and the nodes in other positions within the window are its context nodes. Each center-context pair constitutes a positive training sample. At the same time, other nodes in the corpus are randomly selected as negative examples. During training, the similarity between the center node vector and the context node vector is used to score positive examples, and the similarity between the center node vector and the negative example node vector is used to score negative examples. Through optimization, the score of positive examples is increased while the score of negative examples is decreased. After repeatedly traversing the corpus for several rounds, the vector corresponding to each node is determined, and it is extracted according to the first preset dimension to obtain the spatial flow features of each node. The principle of similarity calculation and optimization is as follows: the similarity between two vectors is measured by the inner product (or normalized inner product). The larger the inner product, the more consistent the direction of the two vectors and the closer they are in space. During optimization, an objective function is constructed with the goal of having a higher similarity between positive examples and negative examples. The two vectors are updated in the direction that improves the objective function. Finally, the vectors of nodes that frequently co-occur in the sequence are made closer to each other, thereby encoding the topological proximity and centrality of the graph into the spatial flow features of a fixed dimension.

[0094] This embodiment employs the aforementioned scheme to first construct a directed graph using entities as nodes and interaction relationships as directed edges in spatial flow feature extraction, generating a random walk sequence on it. This sequence is then fed into a target learning model for training, outputting spatial flow features of a first preset dimension. This transforms the originally non-sequence, difficult-to-model graph structure into node co-occurrence relationships within a sequence. A representation learning model then compresses this co-occurrence statistics into dense, computable, and dimensionally unified node vectors. The preservation of directed edges prevents confusion between active and passive roles in interactions, while biased walks allow features to simultaneously cover both locally adjacent and distant connected structures. This provides a clear and feasible implementation path for spatial flow feature extraction, enabling spatial flow features to realistically and fully characterize the structural position and centrality of nodes in complex topologies. It provides high-quality structural input for subsequent splicing and gating fusion with temporal flow features, thereby supporting the resolution of the spatiotemporal feature fragmentation problem and improving the accuracy of the final core node determination.

[0095] Based on the above implementation scheme, in one feasible implementation, the step of extracting time-flow features from the target dynamic interaction data to obtain time-flow features includes S60~S64: Step S60: Construct a sequence of the target dynamic interaction data based on the interaction time, time interval, interaction type and interaction direction to obtain the entity time sequence.

[0096] It should be noted that interaction time refers to the specific moment when each interaction occurs, and its function is to provide a basis for sorting; time interval refers to the length of time between two adjacent interactions (i.e., the time since the last interaction), and its function is to characterize the frequency of interactions, enabling the sequence model to perceive the non-uniform time characteristics; interaction type refers to the business category identifier to which the interaction belongs, which is discrete category information, and its function is to distinguish different types of interaction behaviors; interaction direction refers to the identifier of whether the entity is the initiator or receiver in this interaction, which is also discrete category information, and its function is to distinguish the role of the entity in the interaction; sequence construction refers to the process of organizing the interactions of the same entity into an ordered sequence in a certain order, and the technology used is time-ordered temporal sequence construction technology; entity temporal sequence refers to the constructed sequence arranged in chronological order with each position carrying the above-mentioned fields, and its function is to serve as the direct input object for time flow feature extraction.

[0097] Understandably, the purpose of this step is to construct a temporally structured input sequence for each network entity that reflects its behavioral evolution. The dynamic characteristics of an entity are reflected in the sequence and spacing of its series of interactive behaviors. If interactions are simply treated as an unordered set for statistical analysis, the trend and rhythm information of behavioral changes will be completely lost. By arranging each interaction from morning to night based on the interaction time, and simultaneously recording the time interval, interaction type, and direction of adjacent interactions, the sequence can carry information such as what happened, when it happened, how long the interval was, what nature it is, and what role it plays. This provides complete material for subsequent step-by-step modeling along the sequence. This allows the dynamic behavior of entities to be expressed as a sequence with a clear sequence and rich fields, thus providing executable input for time-dimensional modeling.

[0098] Specifically, the target dynamic interaction data is first grouped according to entity identifiers. Then, each interaction within a group is sorted from earliest to latest by interaction time. The time interval between each interaction and the previous interaction is calculated, and the interaction type and direction are written into the corresponding positions to obtain the time sequence of the entity. The time interval can be measured in days, hours, or other granularities. Regarding sequence length, excessively long sequences can be truncated to the nearest segment, while excessively short sequences can be padded.

[0099] For example, all interaction records of entity E are retrieved and sorted from earliest to latest by interaction time to obtain an ordered record stream. For each record, the type and direction identifiers of the interaction are written, and the time difference between it and the previous record (i.e., the number of days since the last interaction) is calculated as the time interval at that position. After processing all records in sequence, the temporal sequence of entity E is obtained, where each position in the sequence corresponds to an interaction and carries information such as type, direction, and interval from the previous interaction. The principle of calculating the time interval is as follows: the difference between the interaction times of two adjacent records is used to obtain a value representing the duration difference. This difference directly reflects the density of interaction behavior on the time axis—a small difference indicates dense behavior, and a large difference indicates sparse behavior. Using this difference as a component of the sequence elements allows subsequent models to not only know which interactions occurred, but also how these interactions are distributed over time, thus enabling them to perceive non-uniform spacing characteristics.

[0100] Step S61: Perform linear mapping on the continuous numerical features in the entity time series to obtain linear features.

[0101] It should be noted that continuous numerical features refer to features that take continuous real numbers at each position in the sequence. Typical examples include the logarithm of the interaction weights and the logarithm of the time interval between adjacent interactions. Their role is to characterize the scale and temporal density of the interactions. Linear mapping refers to the process of applying a linear transformation to the input values ​​to obtain a new vector. The technique used is the linear transformation technique of fully connected layers, which is to sum the input values ​​according to the weights and then output the weighted values ​​with a bias. Linear features refer to the vector-form features obtained after the above mapping. Their role is to convert the original low-dimensional, multi-dimensional continuous values ​​into vector expressions with defined dimensions that are compatible with the scale of subsequent discrete features.

[0102] Understandably, the purpose of this step is to convert continuous numerical values ​​at each position in the sequence into a vector form suitable for subsequent processing. Continuous numerical features have very low dimensionality (each item has only one value). If they are directly embedded alongside discrete feature vectors with higher dimensionality, it will cause a serious imbalance in the dimensionality and scale of the components. At the same time, if the original numerical values ​​(especially those after logarithmic transformation) are directly used in subsequent iterative calculations, their scale and distribution may not be suitable for network processing. Through a linear transformation, each continuous numerical value can be mapped to a vector of controllable length, making the continuous and discrete parts more coordinated in terms of dimensionality and numerical scale. At the same time, the linear transformation itself is learnable and can assign appropriate weights to different continuous items during training. Its function is to both upgrade the continuous numerical values ​​to a vector form that is easy to splice and calculate, and obtain a more suitable representation through a learnable transformation, thereby improving the overall representation quality of the sequence elements.

[0103] Specifically, a linear transformation is applied to each continuous numerical feature, converting the value into a vector of fixed length. The vectors of each consecutive feature are then concatenated sequentially to form the linear feature at that position. Alternatively, all continuous numerical features can be first arranged into a single vector, and then subjected to a uniform linear transformation to obtain the linear feature. After the linear transformation, a nonlinear transformation can be superimposed to enhance expressive power, or only the linear transformation can be retained to reduce complexity.

[0104] For example, suppose a sequence contains two consecutive values ​​at a certain position: the logarithm of the interaction weight and the logarithm of the time interval. The system assigns a set of learnable weights and biases to each value: for the logarithm of the interaction weight, it multiplies it by the corresponding weight and adds the bias to obtain a vector of fixed length; the same process is performed on the logarithm of the time interval to obtain another vector; the two vectors are concatenated in order to obtain the linear feature at that position. The calculation principle of the linear mapping is that the transformation formula can be expressed as output = input × weight + bias, which is essentially an affine transformation of the input scalar; when independent weights are assigned to each consecutive value, different consecutive terms are given different scaling and offsets, so that values ​​with different dimensions and value ranges are projected to their respective appropriate representation intervals, achieving both dimensionality increase and making the numerical scale between terms more consistent, thus facilitating their inclusion in the same vector as the embedded vector of discrete features.

[0105] Step S62: Perform word embedding representation on the discrete category features in the entity time sequence to obtain discrete features.

[0106] It should be noted that discrete category features are features with a finite number of discrete category identifiers, typically including interaction type identifiers and interaction direction identifiers. Their purpose is to characterize the nature of the interaction and the role of the entity in the interaction. Word embedding representation is the process of mapping each discrete category identifier to a dense vector. The technique used is the embedding table lookup technique, that is, maintaining a learnable vector for each category identifier and retrieving the corresponding vector as its representation according to the identifier value. Discrete features are vector-form features obtained after the above embedding. Their purpose is to convert category identifiers that cannot be directly calculated and whose size cannot be compared into dense vectors that can participate in distance and similarity calculations, thus unifying them in form with linear features.

[0107] Understandably, the purpose of this step is to convert the category information at each position in the sequence into a computable vector form. The category labels themselves are just symbols, and there is no size or distance relationship between them. If they are directly used in the calculation in the form of numbers, non-existent order and spacing relationships will be artificially introduced (for example, two unrelated categories will be misjudged as similar because their numbers are close). By maintaining an independent and learnable vector for each category label and adjusting these vectors according to the task objective during training, the vectors of categories with similar semantics or functions can be close to each other in space, while the vectors of dissimilar categories can be far apart. This expresses the true relationship between categories in the form of vector distance. Its function is to enable discrete category information to obtain a vector expression that is isomorphic to continuous numerical features, which avoids false order relationships and can be directly concatenated with linear features, so that the information of the sequence elements is complete and uniform in form.

[0108] Specifically, an embedding table is maintained for the interaction type identifier and the interaction direction identifier. The vector is retrieved from the corresponding table based on the identifier value at the current position, and then concatenated sequentially to obtain the discrete feature at that position. Regarding the setting of the embedding dimension, different categories of features can be assigned the same dimension for concatenation, or different dimensions can be set according to the number of values ​​for each category. For the initialization of the embedding table, random initialization followed by updates during training can be used, or task-related statistics can be used for initialization.

[0109] For example, suppose that at a certain position in a sequence, the interaction type identifier is a certain category and the interaction direction identifier is the initiator. The system looks up the corresponding vector in the interaction type embedding table by category value and in the interaction direction embedding table by initiator. Concatenating the two vectors in order yields the discrete feature at that position. The word embedding calculation principle is as follows: the embedding table is mathematically a learnable matrix of the number of categories × vector dimension. The table lookup operation is equivalent to retrieving the row vector using the category number as the row number. During training, the row vectors of this matrix are updated along with the other parameters of the model along with the task objective, so that the row vectors of categories that frequently appear in the same type of samples tend to be similar. Thus, the category identifier, which was originally just a symbol, is transformed into a dense vector with positional and distance meanings in space. It can be directly concatenated with the linear mapping result of continuous features and can also participate in similarity and gating calculations in subsequent iterative calculations.

[0110] Step S63: Concatenate the linear features and discrete features into a comprehensive feature vector, and input the comprehensive feature vector into the gated recurrent unit network based on the time order.

[0111] It should be noted that concatenation refers to the operation of cascading the above two parts into a single vector according to their dimensions; the composite feature vector refers to the complete vector obtained after concatenation, representing a certain interaction in the sequence, and its function is to serve as the input of the recurrent network at the corresponding time step; the temporal order refers to the order in which each interaction in the sequence occurs; the gated recurrent unit network is a recurrent neural network that progressively passes the state along the sequence and controls the degree of retention of historical states and inclusion of new information through a set of learnable gating coefficients. The technique used is a gated recurrent technique with update gates and reset gates, and its function is to model the temporal dependencies in the sequence (especially non-equally spaced sequences).

[0112] Understandably, the purpose of this step is to progressively feed the complete information of each position in the sequence into the recurrent network in chronological order, enabling the network to accumulate and update its understanding of entity behavior along the sequence. Interactions at different positions in the sequence are not independent of each other; early behaviors affect the interpretation of subsequent behaviors. If each position is processed independently, this sequential dependency cannot be utilized. The gated recurrent unit maintains a state that is updated as the sequence progresses and adjusts the update magnitude at each position using a set of gating coefficients (the update gate determines the extent to which the old state is replaced with new information, and the reset gate determines the extent to which the old state is ignored to calculate candidate new states). This allows the network to retain key historical information in long sequences while avoiding the infinite accumulation or excessive forgetting of early information. Since the sequence already carries time interval information, and the gating coefficients themselves are determined by the current input and the historical state, the network can adapt to non-equal interval sequences—the state updates of interactions with longer intervals can be treated differently from those with shorter intervals. Its function is to compress and accumulate the entire interaction sequence of the entity into an internal state that evolves over time, preparing for the next step of taking the final state as a temporal feature.

[0113] Specifically, the linear and discrete features at each position are first concatenated to form a comprehensive feature vector for that position. Then, in ascending order of the sequence, each comprehensive feature vector is sequentially fed into the gate-controlled recurrent unit network as input for each time step, and the network gradually updates its internal state. For processing the sequence direction, forward input from early to late, reverse input from late to early, or a combination of both can be used. Regarding the number of network layers, a single-layer recurrent unit can be used, or multiple layers can be stacked to extract higher-level temporal abstractions.

[0114] For example, for a time series of an entity, the linear and discrete features at each position are first concatenated dimensionally to form a comprehensive feature vector for that position. Then, starting from the earliest interaction, the comprehensive feature vector at that position is fed into the gate control loop unit along with the initial state. Inside the unit, the values ​​of the update gate and the reset gate (coefficients between 0 and 1) are calculated together with the current input and the previous state. Then, the candidate new state is calculated together with the historical state adjusted by the reset gate and the current input. Finally, the update gate performs a weighted mixture between retaining the old state and adopting the candidate new state to obtain the current state. Then, the current state is used as the previous state and the comprehensive feature vector of the next position is used as the current input to repeat the above calculation until the last position. The principle of gating computation is as follows: when the update gate value is close to 1, the current state is almost entirely determined by the candidate new state (i.e., forgetting the old state and adopting new information), and when it is close to 0, the old state is almost completely retained; when the reset gate value is close to 0, the old state is ignored when calculating the candidate new state (i.e., starting from the current input again), and when it is close to 1, the old state is fully utilized; since the values ​​of the two gates are obtained by compressing the current input + the previous state into a range between 0 and 1 after a learnable transformation, the network can decide the magnitude of retention and update for different input content (including different time intervals), thereby achieving differentiated modeling of non-equal interval temporal dependencies.

[0115] Step S64: Extract the hidden state of the gated recurrent unit network at the last moment to obtain the time flow feature of the second preset dimension.

[0116] It should be noted that the final moment refers to the time step corresponding to the last position (i.e., the end of the sequence) of the gated recurrent unit network; the hidden state refers to the internal state vector maintained and passed on by the recurrent unit at each time step, the value of which is determined by the input of that step and all previous steps, and its function is to serve as a compressed memory of the entire sequence before that step; extraction refers to the operation of taking the above state vector out of the network as the output; the second preset dimension refers to the pre-set dimension length of the time flow feature vector, and its function is to make the time flow features and spatial flow features have a definite length that can be spliced ​​and aligned in dimension; the time flow feature refers to the extracted final vector, and its function is to characterize the dynamic evolution law of entity interaction behavior.

[0117] Understandably, the purpose of this step is to transform the states accumulated step by step along the sequence into a fixed-length feature vector representing the entire behavioral evolution of the entity. The hidden state of the recurrent unit at the last time step is obtained after processing all sequence positions step by step. During the calculation process, it absorbs the input information of each position in turn and continuously retains or replaces it through gating. Therefore, it can be regarded as a compressed summary of the entire sequence. The information at the later time step is usually more completely retained, while the key information at the earlier time step can also be continued through selective retention by gating. In contrast, if the state at a certain step in the middle is taken, the behavioral information that occurs afterward will be lost. If the state at all steps is taken, the length is not fixed and is redundant. Its function is to output a temporal flow feature with a fixed dimension, uniform length, and complete evolutionary information, so that it can be directly concatenated with the spatial flow feature with a fixed dimension and enter into subsequent fusion and model training.

[0118] Specifically, the hidden state vector output by the gated recurrent unit network at the last time step is directly taken as the temporal flow feature of the entity. Regarding dimensionality processing, the dimension of the recurrent unit hidden layer can be directly set to the second preset dimension, or it can be mapped to the second preset dimension after extraction via a linear transformation. In multi-layer or bidirectional cases, the state of the last layer at the last time step can be taken, or the states of each direction or each layer at the last time step can be concatenated and then mapped.

[0119] For example, assuming the hidden state dimension of the gated recurrent unit is set to a second preset dimension, after the system has sequentially input all positions of the temporal sequence of an entity, it reads the hidden state vector output at the last time step, which serves as the temporal flow feature of the entity. The same operation is performed on each network entity to obtain the temporal flow feature set of all entities. The calculation principle of the final state, i.e., the sequence summary, is as follows: the update formula of the hidden state at each time step can be understood as new state = retention coefficient × old state + inclusion coefficient × candidate new state. Since each step is based on the old state for weighted updates, the final state is mathematically the result of all historical inputs accumulated through layer-by-layer weighted summation. Therefore, the state at the last time step naturally contains all the information from the beginning to the end of the sequence (decayed according to the gating weight). Taking it as a feature is equivalent to compressing a variable-length behavioral sequence into a fixed-length representation, thus preserving the evolutionary rules and satisfying the requirement of fixed dimensions for subsequent splicing.

[0120] This embodiment employs the aforementioned scheme, first constructing an entity temporal sequence based on interaction time, time interval, interaction type, and interaction direction during temporal feature extraction. Then, it performs linear mapping on continuous numerical features and word embedding on discrete category features, concatenating these two into a comprehensive feature vector and feeding it into a gated recurrent unit network in chronological order. Finally, it extracts the terminal hidden states as temporal features, enabling the sequence to simultaneously carry multi-dimensional information such as behavioral order, temporal density, interaction nature, and role direction. Furthermore, embedding and linear mapping eliminate spurious order relationships in category identifiers and dimensional imbalances in continuous numerical values. By leveraging the selective memorization and updating of non-equidistant temporal dependencies through the gating unit, the variable-length behavioral sequence is compressed into a fixed-length, information-complete evolutionary representation. This provides a clear and feasible implementation path for temporal feature extraction, enabling temporal features to realistically depict the dynamic evolution of entity interaction behavior over time. It provides high-quality temporal-side input for subsequent concatenation and gating fusion with spatial flow features, thereby supporting the resolution of the spatiotemporal feature fragmentation problem and improving the accuracy of core node determination.

[0121] Based on the above implementation scheme, in one feasible implementation, the step of mining and expanding the unlabeled sample set using a positive example unlabeled learning algorithm to obtain an expanded sample set of the unlabeled samples includes S70~S73: Step S70: Select reliable negative samples from the unlabeled sample set.

[0122] It should be noted that the unlabeled sample set refers to a collection of network entity samples whose category labels are unknown. These samples may contain both true core nodes and a large number of non-core nodes, and their purpose is to serve as a candidate sample pool for further exploration. Reliable negative samples refer to samples for which there is sufficient evidence to be certain that they do not belong to the positive class (i.e., they are not core nodes). The technique used is a high-confidence negative class identification technique based on heuristic rules or density clustering. That is, it is determined that the sample does not belong to the positive class based on the degree of difference between the sample and the known positive class in the feature space or based on its own low activity attribute. Filtering refers to the operation of selecting a subset that meets the conditions from all unlabeled samples according to the above criteria.

[0123] Understandably, the purpose of this step is to construct a batch of reliable negative class samples in the absence of any negative class labels. The biggest risk in positive example unlabeled learning is treating all unknown samples as negative classes, thereby incorrectly treating the actual positive classes as negative classes and disrupting the classification boundary. If a batch of samples that are almost impossible to be positive classes can be identified from the unlabeled samples (they are either far from all known positive classes in the feature space or have extremely sparse interactions and lack the scale and connections that core nodes should have), then this batch of samples has high purity as negative classes. Using them to replace all unknown samples in training will not introduce obvious mislabeling. Its role is to make the training of subsequent classifiers based on the reliability of negative classes, reduce the disruption of the decision boundary by mislabeling from the source, and provide a reliable starting point for subsequent iterative expansion.

[0124] Specifically, in one embodiment of this application, a distance-based density clustering approach is adopted. The distance between each unlabeled sample and the known set of positive samples is calculated in the feature space, and the batch of samples furthest apart is selected as reliable negative samples. In another embodiment of this application, a heuristic rule-based approach is adopted. Based on the multi-window sliding characteristics of the samples themselves, samples with very low interaction frequency, small interaction scale, and few connected neighbors (i.e., low activity) are selected as reliable negative samples. Alternatively, the two approaches can be combined: first, low-activity samples are extracted according to rules, and then samples farthest from the positive class are extracted from the remaining samples and merged as reliable negative samples.

[0125] For example, first calculate the distance between the spatiotemporal dual-stream features of each unlabeled sample and the features of each sample in the known positive sample set (e.g., Euclidean distance, which is the square root of the sum of the squared differences of corresponding dimensions in the same vector space). Use the minimum distance from the sample to the positive class set as an indicator of its distance; the larger the indicator, the more the sample deviates from the positive class distribution in the feature space. Sort the samples by this indicator from largest to smallest, and select the top batch of samples as reliable negative samples. At the same time, samples with extremely low interaction frequency and scale in multi-window sliding feature display and very few connected neighbors are also included in the reliable negative samples. The principle of distance calculation is as follows: treat the spatiotemporal dual-stream features of each sample as a point in a high-dimensional space, and the Euclidean distance between two points measures their overall difference in this space. The larger the distance, the more different the sample's overall performance in all dimensions is from the positive class. Using the minimum distance to the positive class set as an indicator, a large indicator means that the sample is also far from the nearest known positive class, so the probability of it belonging to the positive class is very low, and classifying it as a negative class has a high degree of confidence.

[0126] Step S71: Label the known positive samples as the first value and the reliable negative samples as the second value, and train the initial classifier.

[0127] It should be noted that known positive samples refer to samples that have been explicitly labeled as belonging to core nodes, and their role is to provide definite positive class supervision signals; the first value and the second value refer to the label values ​​used to distinguish the categories, usually using the first value as 1 and the second value as 0 (or vice versa) as the category labels for binary classification, transforming the category information into a numerical target that the model can directly fit; the initial classifier refers to the classification model that has not yet been trained on this labeled data and whose parameters need to be optimized, and its role is to serve as the object to be trained in this step, and to have the ability to score unknown samples after training; training refers to the process of using sample features as input and category labels as the target, adjusting the model parameters by optimizing the target, and making the model output gradually approach the label, using supervised binary classification model training techniques.

[0128] Understandably, the purpose of this step is to train a classifier with preliminary discriminative ability using a small but highly pure set of labeled data consisting of a known positive class and a reliable negative class. The decision boundary of the classifier is determined by the distribution of both positive and negative samples during training. If real positive samples are mixed into the negative class, the boundary will be pushed to the wrong position. The negative classes used in this step come from the reliable negative samples selected in step S70, which have an extremely low probability of belonging to the positive class. Therefore, although the training set consisting of known positive samples and reliable negative samples is not large, it has high purity on both the positive and negative sides. The classifier trained based on this can learn a boundary that is relatively close to the truth, rather than a boundary distorted by noisy labels. Its function is to produce an initial scorer that can be used to score the remaining unknown samples, providing a tool for the next iterative expansion.

[0129] Specifically, the labels of known positive samples are assigned as the first value, and the labels of reliable negative samples are assigned as the second value. Using the spatiotemporal dual-stream features corresponding to each sample as input and the labels as the target, a classifier is trained by optimizing the classification loss. For the classifier selection, a decision tree-based ensemble model, a linear model, or a probabilistic model can be used. Regarding the loss handling, equal weighting of the two classes of samples can be applied, or a weighted approach related to the ratio of the number of samples in each class can be used.

[0130] For example, the known core node samples are labeled as 1, and the obtained reliable negative samples are labeled as 0. Their respective spatiotemporal dual-stream feature vectors are extracted to form a training set, and a decision tree ensemble classifier is trained. During training, the classifier outputs a probability estimate that a sample belongs to class 1. This probability and the labeled value are substituted into the loss function to calculate the deviation (the closer the probability is to 1 when labeled as 1, the smaller the deviation; the closer the probability is to 0 when labeled as 0, the smaller the deviation). By adjusting the tree structure and leaf node values ​​in the direction that reduces the total deviation, the classifier's output on the training set gradually approaches the labeled values. The principle of probability estimation and training is as follows: the decision tree ensemble divides the feature space into different leaves through a series of partitions, and uses the proportion of class 1 samples in the leaf as the probability estimate that the region belongs to class 1. The estimates from multiple trees are weighted and summed to obtain a more robust probability. The loss function measures the degree of inconsistency between the estimated probability and the true label. The optimization process continuously seeks a partitioning method that minimizes this inconsistency, thereby gradually bringing the classification boundary closer to the true boundary between positive and negative samples.

[0131] Step S72: Predict the remaining unlabeled samples using the trained initial classifier to obtain the predicted probability of the remaining unlabeled samples.

[0132] It should be noted that the initial classifier after training refers to the classifier that has been trained on known positive samples and reliable negative samples, and whose parameters have been determined. Its role is to serve as a tool for scoring unknown samples. The remaining unlabeled samples refer to the samples remaining in the unlabeled sample set after removing those that have been screened as reliable negative samples. Their role is to serve as the scoring objects in this step and to be the source of potential positive and potential negative samples. Prediction refers to the calculation process of inputting sample features into the classifier and having the classifier give its class tendency. The technique used is forward inference based on the trained model. The predicted probability refers to the probability value output by the classifier that the sample belongs to the positive class (core node). Its role is to express the confidence level of the sample belonging to the positive class in a continuous numerical form, providing a basis for the next step of classification by threshold.

[0133] Understandably, the purpose of this step is to provide a quantifiable and comparable confidence level for samples whose categories are still uncertain, indicating that they belong to the positive class. The resulting classifier has already learned a preliminary boundary between the determined positive class and the credible negative class. By substituting the features of the remaining unknown samples into the model for forward computation, the probability of each sample belonging to the positive class under this boundary can be obtained. Since the probability is a continuous value, it not only gives the class tendency but also the strength of this tendency—a probability close to 1 indicates that the sample clearly falls on the positive side in the feature space with high confidence, while a probability close to 0 indicates that it clearly falls on the negative side with high confidence as well. Samples in the middle range belong to the part of the model that is still uncertain. Its function is to enable subsequent steps to differentiate the remaining samples based on their confidence levels: only high-confidence samples are included in the labeling, while uncertain samples remain in the unknown set for subsequent rounds of judgment, thereby avoiding the introduction of incorrect labeling by forcibly classifying a large number of uncertain samples at once.

[0134] Specifically, the spatiotemporal dual-stream features of the remaining unlabeled samples are input one by one into the trained initial classifier, and the positive class probability output by the classifier is taken as the predicted probability of that sample. Regarding the scoring method, all remaining samples can be scored in a single batch, or they can be scored in batches. As for the utilization of the output, only the probability value itself can be retained for subsequent threshold comparisons, or both the probability value and its corresponding sample identifier can be retained for traceability.

[0135] For example, the spatiotemporal dual-stream feature vectors of the remaining unlabeled samples are organized into a matrix and input into the trained initial classifier all at once. The classifier judges each sample layer by layer according to its learned partitioning rules. Finally, it takes the proportion of Class 1 samples in the leaf region corresponding to each sample and sums them up by weighting multiple trees, outputting a value between 0 and 1, which is the predicted probability that the sample belongs to the core node. The system records the label of each sample and its predicted probability. The principle of probability output calculation is as follows: the classifier divides the feature space into several regions and counts the proportion of positive samples in each region. The region in which a sample falls is used as its historical proportion as its probability estimate. Therefore, the probability essentially reflects what proportion of historical labeled samples with similar features to the sample belong to the positive class. This makes the probability related to both the absolute position of the sample in the feature space and the class composition of the known samples in its neighborhood, thus forming a continuous index that can be used to measure the degree of confidence.

[0136] Step S73: Add samples with predicted probabilities higher than a preset first threshold to the positive sample set as potential positive samples, and add samples with predicted probabilities lower than a preset second threshold to the reliable negative sample set. Iterate through training and prediction until the positive sample set no longer increases significantly or reaches the preset number of iterations to obtain the expanded sample set.

[0137] It should be noted that the preset first threshold refers to a pre-set probability threshold with a relatively high value, which is used to define samples with a sufficiently high predicted probability that can be confidently identified as belonging to the positive class; the preset second threshold refers to another pre-set probability threshold with a relatively low value (below the first threshold), which is used to define samples with a sufficiently low predicted probability that can be confidently identified as belonging to the negative class; potential positive samples refer to samples with a predicted probability higher than the first threshold that are inferred to be positive, and their purpose is to supplement the positive sample set to expand the size of the positive class; the positive sample set and the reliable negative sample set refer to the sets used to store confirmed positive class and confirmed negative class samples, respectively. The roles of both the set and the annotation set are to jointly constitute the labeled data for the current and subsequent training rounds; iterative execution of training and prediction refers to repeatedly executing the process of training a classifier based on the current annotation set, then using the classifier to score the remaining samples and expand the annotation set, using a self-training iterative expansion technique; no longer significantly increasing means that the number of newly added positive samples in a certain round is very small, and the size of the positive sample set tends to be stable; preset iteration number refers to the pre-set upper limit of the number of rounds, and both serve as termination conditions; expanded sample set refers to the labeled sample set obtained after the iteration, which has been expanded on both the positive and negative sides.

[0138] Understandably, the purpose of this step is to expand the training samples by converting high-confidence samples in the unlabeled samples into usable labels through a cycle of scoring, high-confidence expansion, retraining, and re-scoring, based on the initial classifier. The initial classifier is trained only on a small set of confirmed positive and credible negative classes, resulting in limited generalization ability. However, its judgments on samples clearly biased to one side (samples with very high or very low prediction probabilities) are usually quite reliable. Adding these high-confidence samples to their corresponding categories increases the size and coverage of the training samples. Retraining with a larger training set yields a more robust classifier, which in turn can better cover more... Samples initially in the uncertainty range are distinguished with higher confidence, and this process is repeated to gradually expand the annotation. Only samples above the first threshold or below the second threshold are included, while samples between the two thresholds remain in the unknown set, ensuring that the annotations added in each round maintain high purity and avoiding noise accumulation. The process stops when the size of the positive sample set no longer increases significantly (indicating that the positive classes that can be mined with high confidence have been basically exhausted) or when the upper limit of the round is reached, preventing meaningless infinite loops. Its function is to enable the training samples to expand in size while maintaining the credibility of the annotations, thereby alleviating overfitting and boundary shift caused by the scarcity of positive samples.

[0139] Specifically, in each round, samples with predicted probabilities higher than a first threshold are added to the positive sample set, and samples with probabilities lower than a second threshold are added to the reliable negative sample set. The classifier is then retrained based on the expanded sets, and the new classifier is used to re-predict the remaining samples after the previous round of expansion, proceeding to the next round. Regarding threshold settings, the first threshold can be set to a relatively high value close to 1, and the second threshold can be set to a relatively low value close to 0 to strictly ensure the purity of the included samples; alternatively, the thresholds can be gradually relaxed during iteration to accelerate the expansion speed. Termination can be determined when the number of newly added positive samples in a round falls below a certain scale, indicating that there is no longer a significant increase, or it can terminate when a pre-set round limit is reached.

[0140] For example, let the first threshold be a relatively high value close to 1 and the second threshold be a relatively low value close to 0. In the first round, the predicted probabilities are compared with the two thresholds one by one: if the predicted probability of a sample is higher than the first threshold, it is added to the positive sample set; if the predicted probability of another sample is lower than the second threshold, it is added to the reliable negative sample set; samples with predicted probabilities between the two thresholds are left unprocessed and remain in the unknown set. Subsequently, the system retrains the classifier with the expanded positive sample set and reliable negative sample set, and then uses the new classifier to re-predict the remaining unknown samples and expands the set again according to the thresholds, repeating this process. If the number of newly added positive sample sets in a certain round is very small, or the number of rounds executed has reached the upper limit, the iteration stops, and the current positive sample set and reliable negative sample set are output together as the expanded sample set. The computational principle of iterative improvement is as follows: each round of expansion increases the training set on both the positive and negative sides simultaneously, and the confidence of newly added samples is guaranteed by a threshold. Therefore, the boundary fitted by the new classifier is obtained on a larger scale and with wider coverage of labeled data, and its ability to distinguish samples in the middle uncertain interval is enhanced, so that more samples can be distinguished with high confidence in the next round. The setting of two thresholds ensures that each round only absorbs the two ends of the samples that the model is most confident in, leaving the uncertain samples to be processed by the subsequent more robust classifier, thereby suppressing the accumulation of mislabeled samples while expanding the scale.

[0141] This embodiment employs the above-described scheme: first, reliable negative samples are selected from the unlabeled sample set; then, an initial classifier is trained using known positive samples and reliable negative samples labeled with two different values; subsequently, the classifier is used to predict the probability of the remaining unlabeled samples; finally, high-confidence samples are included in both the positive and negative sides using two thresholds, and training and prediction are iteratively performed until convergence. This ensures that the negative side is no longer composed entirely of unknown samples, but rather of high-purity negative samples identified based on their distributional distance or low activity attributes. This avoids misclassifying true positive samples from unknown samples as negative samples, thus disrupting the classification boundary. Simultaneously, the dual thresholds only absorb the samples at the two ends that the model is most confident in, leaving uncertain samples for subsequent rounds. This ensures that each round of expansion increases the training scale while maintaining the reliability of the labels, and through iteration, the classifier is gradually enhanced, distinguishing more previously uncertain samples with higher confidence. Therefore, this provides a clear, convergent, and purity-controllable two-step implementation path for positive example unlabeled learning. It can robustly construct a sufficiently large and reliably labeled extended sample set in scenarios where positive samples are extremely scarce and unlabeled samples are massive, significantly reducing the risk of model overfitting and decision boundary shift, and directly improving the technical problem of classifier failure caused by the scarcity of positive samples.

[0142] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the dynamic network core node prediction method of this application. Any simple modifications based on this technical concept are within the protection scope of this application.

[0143] This application also provides a dynamic network core node prediction device, please refer to... Figure 2 The dynamic network core node prediction device includes: Feature extraction module 200 is used to extract and fuse spatiotemporal dual-stream data of the target dynamic interaction data to obtain spatiotemporal dual-stream features of the target dynamic interaction data; The sample expansion module 201 is used to mine and expand the unlabeled sample set through a positive example unlabeled learning algorithm to obtain an expanded sample set of the unlabeled samples; Model training module 202 is used to train the initial core node prediction model based on the spatiotemporal dual-stream features and the extended sample set to obtain the target core node prediction model. The node prediction module 203 is used to perform inference prediction on the entity data to be predicted in the dynamic network through the target core node prediction model to obtain the core nodes in the dynamic network.

[0144] The dynamic network core node prediction device provided in this application, employing the dynamic network core node prediction method in the above embodiments, can solve the technical problem of insufficient prediction accuracy of core nodes in dynamic networks. Compared with the prior art, the beneficial effects of the dynamic network core node prediction device provided in this application are the same as those of the dynamic network core node prediction method provided in the above embodiments, and other technical features in the dynamic network core node prediction device are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0145] This application provides a dynamic network core node prediction device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the dynamic network core node prediction method in Embodiment 1 above.

[0146] The following is for reference. Figure 3 This document illustrates a structural schematic diagram of a dynamic network core node prediction device suitable for implementing embodiments of this application. The dynamic network core node prediction device in this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 3 The dynamic network core node prediction device shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0147] like Figure 3As shown, the dynamic network core node prediction device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory 1002 or a program loaded from a storage device 1003 into a random access memory 1004. The random access memory 1004 also stores various programs and data required for the operation of the dynamic network core node prediction device. The processing unit 1001, the read-only memory 1002, and the random access memory 1004 are interconnected via a bus 1005. An input / output interface 1006 is also connected to the bus. Typically, the following systems can be connected to the input / output interface 1006: input devices 1007 including, for example, a touch screen, touchpad, keyboard, mouse, image sensor, microphone, accelerometer, gyroscope, etc.; output devices 1008 including, for example, a liquid crystal display (LCD), speaker, vibrator, etc.; storage devices 1003 including, for example, magnetic tape, hard disk, etc.; and communication devices 1009. The communication device 1009 allows the dynamic network core node prediction device to communicate wirelessly or wiredly with other devices to exchange data. Although the figure shows dynamic network core node prediction devices with various systems, it should be understood that implementation or possession of all the systems shown is not required. More or fewer systems may be implemented alternatively.

[0148] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from read-only memory 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0149] The dynamic network core node prediction device provided in this application, employing the dynamic network core node prediction method in the above embodiments, can solve the technical problem of insufficient prediction accuracy of core nodes in dynamic networks. Compared with the prior art, the beneficial effects of the dynamic network core node prediction device provided in this application are the same as those of the dynamic network core node prediction method provided in the above embodiments, and other technical features in this dynamic network core node prediction device are the same as those disclosed in the method of the previous embodiment, and will not be repeated here.

[0150] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0151] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0152] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the dynamic network core node prediction method in the above embodiments.

[0153] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems or devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0154] The aforementioned computer-readable storage medium may be included in the dynamic network core node prediction device; or it may exist independently and not be assembled into the dynamic network core node prediction device.

[0155] The aforementioned computer-readable storage medium carries one or more programs. When these programs are executed by the dynamic network core node prediction device, the dynamic network core node prediction device performs spatiotemporal dual-stream extraction and fusion on the target dynamic interaction data to obtain the spatiotemporal dual-stream features of the target dynamic interaction data; mines and expands the unlabeled sample set using a positive example unlabeled learning algorithm to obtain an expanded sample set of the unlabeled samples; trains an initial core node prediction model based on the spatiotemporal dual-stream features and the expanded sample set to obtain a target core node prediction model; and performs inference prediction on the entity data to be predicted in the dynamic network using the target core node prediction model to obtain the core nodes in the dynamic network.

[0156] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0157] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0158] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0159] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described dynamic network core node prediction method, thereby solving the technical problem of insufficient prediction accuracy for core nodes in dynamic networks. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the dynamic network core node prediction method provided in the above embodiments, and will not be repeated here.

[0160] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the dynamic network core node prediction method described above.

[0161] The computer program product provided in this application can solve the technical problem of insufficient prediction accuracy of core nodes in dynamic networks. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the dynamic network core node prediction method provided in the above embodiments, and will not be repeated here.

[0162] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A method for predicting core nodes in a dynamic network, characterized in that, The dynamic network core node prediction method includes: Spatiotemporal dual-stream extraction and fusion of the target dynamic interaction data are performed to obtain the spatiotemporal dual-stream features of the target dynamic interaction data; The unlabeled sample set is expanded by mining the unlabeled sample set through a positive example unlabeled learning algorithm to obtain the expanded sample set of the unlabeled samples; The initial core node prediction model is trained based on the spatiotemporal dual-stream features and the extended sample set to obtain the target core node prediction model. The core nodes in the dynamic network are obtained by reasoning and predicting the entity data to be predicted in the dynamic network through the target core node prediction model.

2. The dynamic network core node prediction method as described in claim 1, characterized in that, Before the step of extracting and fusing spatiotemporal dual-stream data of the target dynamic interaction data to obtain the spatiotemporal dual-stream features of the target dynamic interaction data, the method further includes: Obtain the original dynamic interaction data of network entities, and perform data cleaning processing on the original dynamic interaction data; A multi-window sliding feature is constructed based on the cleaned original dynamic interaction data. The interaction weights of the multi-window sliding feature are then subjected to a logarithmic transformation to obtain the target dynamic interaction data.

3. The dynamic network core node prediction method as described in claim 2, characterized in that, The steps for constructing multi-window sliding features based on the cleaned original dynamic interaction data include: Several time windows of different target lengths are set, and historical backtracking is performed based on the current time node to obtain the interaction frequency, total interaction weight and number of interacting neighbor nodes of network entities within the corresponding time window. The interaction frequency, total interaction weight, and number of interacting neighbor nodes obtained under the time window are combined to obtain the multi-window sliding feature corresponding to the time window.

4. The dynamic network core node prediction method as described in claim 1, characterized in that, The step of extracting and fusing spatiotemporal dual-stream data of the target dynamic interaction data to obtain the spatiotemporal dual-stream features of the target dynamic interaction data includes: Spatial flow features are extracted from the target dynamic interactive data to obtain spatial flow features; Temporal flow features are extracted from the target dynamic interactive data to obtain temporal flow features; The spatial flow features and the temporal flow features are concatenated to obtain the initial concatenated features; The initial spliced ​​features are input into an adaptive gated fusion network for cross-modal feature interaction and deep fusion to obtain the spatiotemporal dual-stream features.

5. The dynamic network core node prediction method as described in claim 4, characterized in that, The step of extracting spatial flow features from the target dynamic interactive data to obtain spatial flow features includes: A directed graph is constructed using network entities as nodes and interaction relationships as directed edges, and a random walk sequence is generated on the directed graph; The random walk sequence is input into the target learning model for training, and the spatial flow features of the first preset dimension are output.

6. The dynamic network core node prediction method as described in claim 4, characterized in that, The step of extracting time-flow features from the target dynamic interaction data to obtain the time-flow features includes: Based on the interaction time, time interval, interaction type and interaction direction, the target dynamic interaction data is sequenced to obtain the entity time sequence. Linear features are obtained by linearly mapping the continuous numerical features in the time series of the entity; The discrete category features in the entity time sequence are represented by word embedding to obtain discrete features; The linear and discrete features are concatenated into a comprehensive feature vector, which is then input into a gated recurrent unit network based on time order. Extract the hidden state of the gated recurrent unit network at the last moment to obtain the time flow feature of the second preset dimension.

7. The dynamic network core node prediction method as described in claim 1, characterized in that, The step of mining and expanding the unlabeled sample set using a positive example unlabeled learning algorithm to obtain an expanded sample set of the unlabeled samples includes: Reliable negative samples are selected from the unlabeled sample set; The known positive samples are labeled with the first value, and the reliable negative samples are labeled with the second value. The initial classifier is then trained. The remaining unlabeled samples are predicted by the trained initial classifier to obtain the predicted probability of the remaining unlabeled samples; Samples with predicted probabilities higher than a preset first threshold are added to the positive sample set as potential positive samples, and samples with predicted probabilities lower than a preset second threshold are added to the reliable negative sample set. Training and prediction are performed iteratively until the positive sample set no longer increases significantly or reaches the preset number of iterations, thus obtaining the expanded sample set.

8. A dynamic network core node prediction device, characterized in that, The dynamic network core node prediction device includes: The feature extraction module is used to extract and fuse the target dynamic interaction data in both spatiotemporal streams to obtain the spatiotemporal dual-stream features of the target dynamic interaction data. The sample expansion module is used to mine and expand the unlabeled sample set through a positive example unlabeled learning algorithm to obtain an expanded sample set of the unlabeled samples; The model training module is used to train the initial core node prediction model based on the spatiotemporal dual-stream features and the extended sample set to obtain the target core node prediction model. The node prediction module is used to perform inference and prediction on the entity data to be predicted in the dynamic network through the target core node prediction model, so as to obtain the core nodes in the dynamic network.

9. A dynamic network core node prediction device, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the dynamic network core node prediction method as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the dynamic network core node prediction method as described in any one of claims 1 to 7.

Images