Hidden individual-oriented structure prediction method and device and computer equipment

Abstract

The invention relates to a hidden individual-oriented structure prediction method and device and computer equipment. The method comprises the following steps: acquiring a state track of a visible individual in a target system; performing hidden trajectory prediction on the state trajectory of the visible individual through a first hidden state prediction network to obtain an initial prediction hidden state trajectory; splicing the initial prediction hidden state trajectory with the state trajectory of the visible individual to obtain a first spliced state trajectory, and performing system structure prediction on the first spliced state trajectory through a target structure inference network to obtain target prediction structure features; and performing hidden trajectory prediction on the target prediction structure feature and the state trajectory of the visible individual through a second hidden state prediction network to obtain a target prediction hidden state trajectory. By adopting the method, the structure inference effect can be improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method, apparatus, computer device, computer-readable storage medium, and computer program product for predicting the structure of hidden individuals. Background Technology

[0002] In scenarios such as physical systems, biological systems, social networks, and multi-entity control, systems are typically composed of multiple interacting individuals and can be described using a graph structure. Nodes represent individuals, and edges represent interactions or causal relationships between them. An accurate understanding of this graph structure is fundamental for system analysis, anomaly detection, and control decisions. However, in many real-world scenarios, due to limitations in sensor coverage, sampling noise, cost constraints, and privacy protection, some individuals in the system cannot be directly observed. Their existence and state affect system dynamics but do not appear in the observational data; these are known as "hidden individuals." For example, in ecosystems, some species may go unmonitored; in molecular biology, certain intermediate reactants or proteins are difficult to measure; and in social networks, some accounts are hidden or anonymous. In these cases, traditional structural inference methods that assume "all individuals are observable" face severe performance degradation. Structural inference technology aims to automatically reconstruct the underlying interaction graph structure using only the state trajectory data of individuals evolving over time, and then predict future trajectories based on this structure.

[0003] In traditional techniques, unsupervised learning based on latent graph structures is used to predict future states. When all individual states are observable, the structure inference works well, but when some individual states are not observable, the structure inference works poorly. Summary of the Invention

[0004] Therefore, it is necessary to provide a method, apparatus, computer device, computer-readable storage medium, and computer program product for predicting the structure of hidden individuals that can improve the structure inference effect, in order to address the above-mentioned technical problems.

[0005] Firstly, this application provides a structural prediction method for hidden individuals, including:

[0006] Obtain the state trajectories of visible individuals in the target system;

[0007] The state trajectory of the visible individual is predicted by the first hidden state prediction network to obtain the initial predicted hidden state trajectory.

[0008] The initial predicted hidden state trajectory is concatenated with the state trajectory of the visible individual to obtain the first concatenated state trajectory. The target structure inference network is then used to predict the system structure of the first concatenated state trajectory to obtain the target predicted structure features.

[0009] The hidden trajectory prediction of the target is obtained by using a second hidden state prediction network to predict the hidden trajectory of the target prediction structure features and the state trajectory of the visible individuals.

[0010] Secondly, this application also provides a structural prediction device for hidden individuals, comprising:

[0011] The visible trajectory acquisition module is used to acquire the state trajectories of visible individuals in the target system;

[0012] The initial hidden state prediction module is used to predict the hidden trajectory of the visible individual's state trajectory through the first hidden state prediction network to obtain the initial predicted hidden state trajectory.

[0013] The system structure prediction module is used to concatenate the initial predicted hidden state trajectory with the state trajectory of the visible individual to obtain a first concatenated state trajectory, and to perform system structure prediction on the first concatenated state trajectory through the target structure inference network to obtain the target predicted structure features.

[0014] The target hidden state prediction module uses a second hidden state prediction network to predict the hidden trajectory of the target prediction structure features and the state trajectory of the visible individuals, thereby obtaining the target prediction hidden state trajectory.

[0015] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the structural prediction method for hidden individuals provided in the first aspect.

[0016] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the structure prediction method for hidden individuals provided in the first aspect.

[0017] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the structure prediction method for hidden individuals provided in the first aspect.

[0018] The aforementioned method, apparatus, computer device, computer-readable storage medium, and computer program product for structure prediction of hidden individuals, through a first hidden state prediction network, predicts the hidden trajectory of the visible individual's state trajectory to obtain an initial predicted hidden state trajectory. This initial predicted hidden state trajectory is then concatenated with the visible individual's state trajectory to obtain a first concatenated state trajectory. A target structure inference network is then used to predict the system structure of the first concatenated state trajectory to obtain the target predicted structural features. Finally, a second hidden state prediction network is used to predict the hidden trajectory of both the target predicted structural features and the visible individual's state trajectory to obtain the target predicted hidden state trajectory. This approach enables a closed loop between hidden state prediction and structure inference, simultaneously obtaining more accurate hidden state estimation and structure inference results even in the presence of hidden individuals and some observable noise, thus improving the accuracy of structure inference. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a diagram illustrating the application environment of a structure prediction method for hidden individuals in one embodiment.

[0021] Figure 2 This is a flowchart illustrating a structure prediction method for hidden individuals in one embodiment;

[0022] Figure 3 This is a flowchart illustrating structural constraints using a multi-head attention mechanism in one embodiment.

[0023] Figure 4 This is a flowchart illustrating a structure prediction method for hidden individuals in another embodiment;

[0024] Figure 5 This is a structural block diagram of a structure prediction device for hidden individuals in one embodiment;

[0025] Figure 6 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0027] It should be noted that the terms "first," "second," etc., used in this application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more. The term "and / or" used in this application refers to one of the embodiments, or any combination of multiple embodiments.

[0028] The structure prediction method for hidden individuals provided in this application can be applied to, for example... Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104 or placed on a cloud or other network server. Server 104 can obtain the state trajectories of visible individuals in the target system sent by terminal 102. Server 104 uses a first hidden state prediction network to predict the hidden trajectory of the visible individuals' state trajectories, obtaining an initial predicted hidden state trajectory. This initial predicted hidden state trajectory is then concatenated with the state trajectories of the visible individuals to obtain a first concatenated state trajectory. A target structure inference network is then used to predict the system structure of the first concatenated state trajectory, obtaining the target predicted structural features. Finally, a second hidden state prediction network is used to predict the hidden trajectory of both the target predicted structural features and the state trajectories of the visible individuals, obtaining the target predicted hidden state trajectory. Server 104 can then return the obtained target predicted hidden state trajectory to terminal 102. The terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, drones, low-altitude aircraft, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, smart in-vehicle devices, and projection devices. Portable wearable devices can include smartwatches, smart bracelets, and head-mounted displays. Head-mounted displays can be virtual reality (VR) devices, augmented reality (AR) devices, and smart glasses. The server 104 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services. It should be noted that the structure prediction method for hidden individuals provided in this application embodiment is applicable not only to the above-mentioned server-terminal interaction application scenarios but also to application scenarios involving a single server, a single terminal, server-to-server interaction, or terminal-to-terminal interaction.

[0029] In one exemplary embodiment, such as Figure 2As shown, a structure prediction method for hidden individuals is provided, which can be applied to... Figure 1 Taking the server in the example, the explanation includes the following steps 202 to 208. Wherein:

[0030] Step 202: Obtain the state trajectories of visible individuals in the target system.

[0031] The target system can be any system comprising visible and hidden individual states. Visible individuals are those whose state trajectories can be directly observed. Hidden individuals are those whose state trajectories cannot be directly observed; they objectively exist within the target system and dynamically influence its dynamics, but their state trajectories cannot be directly observed or collected. An individual refers to a basic entity within a target system that has independent states and participates in interactions; such as physical particles, robots, user nodes, or molecules. State trajectories are used to characterize the state development trend of an individual. For example, in an ecosystem, some species can be monitored, while others cannot be directly monitored. In social networks, some accounts are hidden or anonymous and not visible. In molecular biology, certain intermediate reactants or proteins are difficult to measure, etc. In practical applications, it is easy to obtain the state trajectories of visible individuals, and then predict the state trajectories of hidden individuals based on these trajectories. After obtaining the state trajectories of both visible and hidden individuals, the interaction structure of the system composed of these individuals is easily obtained.

[0032] In practical applications, the state trajectory of visible individuals in the target system can be obtained through the server's human-computer interaction interface or input data interface.

[0033] Step 204: The hidden trajectory of the visible individual is predicted by the first hidden state prediction network to obtain the initial predicted hidden state trajectory.

[0034] The first hidden state prediction network is used to predict hidden trajectories based on the state trajectories of visible individuals. For example, by inputting the state trajectories of visible individuals in the target system into the first hidden state prediction network, the state trajectories of hidden individuals in the same target system can be predicted. The initial predicted hidden state trajectory refers to the state trajectory of the hidden individuals predicted by the first hidden state prediction network based on the state trajectories of visible individuals. The initial hidden state trajectory may include the state trajectories of one or more hidden individuals, and the number of hidden individuals can be determined according to the actual application scenario.

[0035] For example, the first hidden state prediction network can predict hidden trajectories based on a Set Transformer architecture, DeepSets, or graph attention networks. In other words, the first hidden state prediction network can be trained based on a Set Transformer architecture, DeepSets, or graph attention networks. It should be noted that the first hidden state prediction network does not incorporate the structure of the target system into the prediction process; it only predicts the hidden trajectory based on the state trajectory of the visible individuals.

[0036] In an exemplary embodiment, a first prediction network can be trained using state trajectory samples of visible individuals and state trajectory samples of hidden individuals. That is, the state trajectory of hidden individuals is predicted using the state trajectory of visible individuals until a corresponding first training condition is met, thus obtaining a first hidden state prediction network. The first prediction network can, for example, be a prediction network based on a Set Transformer architecture.

[0037] Step 206: The initial predicted hidden state trajectory is concatenated with the state trajectory of the visible individuals to obtain the first concatenated state trajectory. The target structure inference network is then used to predict the system structure of the first concatenated state trajectory to obtain the target predicted structure features.

[0038] Here, the first spliced ​​state trajectory refers to the state trajectory obtained by splicing the initial predicted hidden state trajectory with the state trajectories of visible individuals. The target predicted structural features refer to the structural features predicted by the target structure inference network on the first spliced ​​state trajectory. The target structure inference network is a network capable of predicting the system structure based on individual state trajectories. The target structure inference network can be an existing structure inference network or a trained structure inference network. For example, it can be trained based on models such as NRI (Neural Relational Inference), DNRI (Dynamic NRI), FNRI (Factorized NRI), GNP (Graph Neural Process), and VGAE (Variational Graph Auto-Encoders).

[0039] In practical applications, the server can concatenate the obtained initial predicted hidden state trajectory with the state trajectory of the visible individuals in terms of dimensions to obtain the first concatenated state trajectory. Then, the target structure inference network can be used to predict the system structure of the first concatenated state trajectory to obtain the target prediction result features.

[0040] For example, the target structure inference network can be a trained structure inference network. For instance, the predicted state trajectory of a hidden individual can be concatenated with the state trajectory of a visible individual to obtain a concatenated state trajectory. The initial structure inference network can then predict the result of the concatenated state trajectory to obtain predicted structural features. Based on the difference between the predicted structural features and the true structural features, the initial structure inference network can be trained until a second training condition is met to obtain the target structure inference network.

[0041] Step 208: The hidden trajectory prediction of the target is obtained by using the second hidden state prediction network to predict the target prediction structural features and the state trajectories of visible individuals.

[0042] The second hidden state prediction network is used to predict hidden trajectories based on the target prediction structure features and the state trajectories of visible individuals. The target prediction hidden state trajectory refers to the hidden individual's state trajectory predicted by the second hidden state prediction network based on the target prediction structure features and the state trajectories of visible individuals. It is easy to understand that, compared to the first hidden state prediction network, the second hidden state prediction network introduces target prediction structure features for hidden trajectory prediction, resulting in more accurate predictions.

[0043] In practical applications, the target prediction structural features corresponding to the target system and the state trajectories of visible individuals can be input into the second hidden state prediction network to obtain the target prediction hidden state trajectories corresponding to the hidden individuals in the target system. The second hidden state prediction network can be implemented using the Set Transformer architecture, just like the first hidden state prediction network. However, a structural mask is explicitly introduced in the attention calculation of the encoder and decoder of the second hidden state prediction network to constrain the system structure.

[0044] In one example, the second prediction network can be trained based on the second predicted structure features predicted by the target structure inference network and the state trajectory samples of visible individuals until the difference between the output state trajectory of the hidden individual and the state trajectory samples of the hidden individual satisfies the third training condition, thus obtaining the second hidden state prediction network. The second prediction network can be the same as or different from the first prediction network.

[0045] The aforementioned structure prediction method for hidden individuals uses a first hidden state prediction network to predict the hidden trajectory of the visible individuals' state trajectories, obtaining an initial predicted hidden state trajectory. This initial predicted hidden state trajectory is then concatenated with the visible individuals' state trajectories to obtain a first concatenated state trajectory. A target structure inference network then uses this first concatenated state trajectory to predict the system structure, obtaining the target predicted structural features. Finally, a second hidden state prediction network uses the target predicted structural features and the visible individuals' state trajectories to predict the hidden trajectory, obtaining the target predicted hidden state trajectory. This method achieves a closed loop between hidden state prediction and structure inference, simultaneously obtaining more accurate hidden state estimation and structure inference results even in the presence of hidden individuals and some observable noise, thus improving the accuracy of structure inference.

[0046] In some embodiments, before predicting the hidden trajectory of the visible individual's state trajectory using the first hidden state prediction network in step 204 to obtain the initial predicted hidden state trajectory, the above method further includes:

[0047] A training dataset is obtained, which includes state trajectory samples of multiple individuals. A first number of individuals in the training dataset are labeled as visible individuals, and a second number of individuals other than the first number are labeled as hidden individuals. The state trajectory samples of the visible individuals are used to predict the hidden trajectories through a first prediction network to obtain the first predicted hidden state trajectory. The parameters of the first prediction network are adjusted according to the difference between the first predicted hidden state trajectory and the state trajectory samples of the hidden individuals until the first training condition is met, thus obtaining the first hidden state prediction network.

[0048] In this method, the state trajectory samples of multiple individuals are used as the training dataset to train the first hidden state prediction network. A first number of individuals in the training dataset are labeled as visible individuals, and a second number of individuals are labeled as hidden individuals. That is, the training dataset includes the state trajectory samples of the sum of the first and second numbers of individuals. It is easy to understand that the first or second number can be set according to the actual application scenario.

[0049] In practical applications, before predicting the hidden trajectory of a visible individual's state trajectory using a first hidden state prediction network to obtain the initial predicted hidden state trajectory, it is necessary to acquire the first hidden state prediction network. This first hidden state prediction network can be trained. The training method for the first hidden state prediction network may include: predicting the hidden trajectory of visible individual state trajectory samples using the first prediction network to obtain the first predicted hidden state trajectory; adjusting the parameters of the first prediction network based on the difference between the first predicted hidden state trajectory and the state trajectory samples labeled as hidden individuals, until a first training condition is met, thus obtaining the first hidden state prediction network. The first training condition is used to characterize the convergence of the training of the first hidden state prediction network. The first training condition may be, for example, that the difference between the first predicted hidden state trajectory and the state trajectory samples labeled as hidden individuals is less than a first difference threshold, or that the number of training iterations reaches a first quantity threshold, etc.

[0050] It is easy to understand that the first predicted hidden state trajectory should include the predicted trajectories of the second number of hidden individuals. When calculating the difference between the first predicted hidden state trajectory and the state trajectory samples labeled as hidden individuals, the difference between each pair of matched first predicted hidden state trajectories and the state trajectory samples of the corresponding hidden individuals is calculated separately. For example, the MSE between each matched first predicted hidden state trajectory and the state trajectory samples of the corresponding hidden individuals is used as the corresponding training loss. The parameters of the first prediction network are adjusted according to the training loss until the training loss is less than the preset loss threshold, the first prediction network converges, and the first hidden state prediction network is obtained.

[0051] In this embodiment, a first prediction network is used to predict the hidden trajectory of the state trajectory samples of visible individuals to obtain a first predicted hidden state trajectory. Based on the difference between the first predicted hidden state trajectory and the state trajectory samples of hidden individuals, the parameters of the first prediction network are adjusted until the first training condition is met to obtain the first hidden state prediction network. This enables the accurate training of the first hidden state prediction network based on the state trajectory samples of visible individuals and the state trajectory samples of hidden individuals.

[0052] In some embodiments, before performing system structure prediction on the first spliced ​​state trajectory through the target structure inference network in step 206 to obtain the target predicted structure features, the above method further includes:

[0053] A training dataset is obtained, which includes state trajectory samples of multiple individuals. A first number of individuals in the training dataset are labeled as visible individuals, and a second number of individuals are labeled as hidden individuals. A first hidden state prediction network is used to predict the hidden trajectories of the visible individuals' state trajectory samples to obtain a second predicted hidden state trajectory. The second predicted hidden state trajectory is concatenated with the state trajectory samples of the visible individuals to obtain a second concatenated state trajectory. An initial structure inference network is used to predict the system structure of the second concatenated state trajectory to obtain a first predicted structure feature. Based on the difference between the first predicted structure feature and the corresponding reference structure feature in the training dataset, the parameters of the initial structure inference network are adjusted until a second training condition is met to obtain the target structure inference network.

[0054] It is easy to understand that the training dataset can be the same as the training dataset used to train the first hidden state prediction network. In practical applications, before using the target structure inference network to predict the system structure of the first spliced ​​state trajectory and obtain the target predicted structure features, it is necessary to determine the target structure inference network.

[0055] For example, the target structure inference network can be trained. The training method for the target structure inference network may include: the server can use a trained first hidden state prediction network to predict the hidden trajectories of visible individual state trajectory samples to obtain a second predicted hidden state trajectory; then, the second predicted hidden state trajectory is concatenated with the visible individual state trajectory samples to obtain a second concatenated state trajectory; the second concatenated state trajectory is input into the initial structure inference network to obtain a first predicted structural feature; the difference between the first predicted structural feature and the reference structural feature corresponding to the training dataset is determined; the parameters of the initial structure inference network are adjusted based on this difference; and the above training process is iterated until a second training condition is met to obtain the target structure inference network. The second training condition is used to characterize the convergence of the target structure inference network training. The second training condition may be, for example, that the difference between the first predicted structural feature and the reference structural feature corresponding to the training dataset is less than a second difference threshold, or that the number of training iterations reaches a second quantity threshold, etc.

[0056] In one example, the second predicted hidden state trajectory is concatenated with the state trajectory samples of visible individuals to obtain the second concatenated state trajectory. This second concatenated state trajectory is then input into the initial structure inference network to obtain the first predicted structural features of the corresponding system and the future state trajectories of the visible individuals. The first predicted structural features characterize the predicted interaction structure between individuals in the target system. The future state trajectory refers to the state trajectory of the visible individuals in the future; easily understood, the aforementioned state trajectory samples of the visible individuals are equivalent to the historical state trajectories of the visible individuals.

[0057] In this embodiment, a first hidden state prediction network is used to predict the hidden trajectory of the state trajectory samples of visible individuals to obtain a second predicted hidden state trajectory. The second predicted hidden state trajectory is then concatenated with the state trajectory samples of visible individuals to obtain a second concatenated state trajectory. An initial structure inference network is used to predict the system structure of the second concatenated state trajectory to obtain a first predicted structure feature. Based on the difference between the first predicted structure feature and the reference structure feature corresponding to the training dataset, the parameters of the initial structure inference network are adjusted until the second training condition is met, thus obtaining a target structure inference network with high accuracy.

[0058] In some embodiments, before step 208, which uses a second hidden state prediction network to predict the hidden trajectory of the target prediction structure features and the state trajectories of visible individuals to obtain the target prediction hidden state trajectory, the above method further includes:

[0059] The target structure inference network predicts the system structure of the second spliced ​​state trajectory to obtain the second predicted structure features. The second prediction network then predicts the hidden trajectory of the second predicted structure features and the state trajectory samples of visible individuals to obtain the third predicted hidden state trajectory. Based on the difference between the third predicted hidden state trajectory and the state trajectory samples of hidden individuals, the parameters of the second prediction network are adjusted until the third training condition is met, thus obtaining the second hidden state prediction network.

[0060] Among them, the target structure inference network is the trained structure inference network. The second prediction network is the hidden trajectory prediction network before training, and the second hidden state prediction network is the hidden trajectory prediction network after training. Before using the second hidden state prediction network to predict the hidden trajectory of the target predicted structure features and the state trajectories of visible individuals to obtain the target predicted hidden state trajectory, the second hidden state prediction network needs to be determined first.

[0061] In practical applications, the second hidden state prediction network can be obtained through training. For example, the training method for the second hidden state prediction network includes: using a pre-trained target structure inference network to predict the system structure of the second spliced ​​state trajectory, obtaining the second predicted structural features; using the second prediction network to predict the hidden trajectory of the second predicted structural features and the state trajectory samples of visible individuals, obtaining the third predicted hidden state trajectory; and adjusting the parameters of the second prediction network based on the difference between the third predicted hidden state trajectory and the state trajectory samples of hidden individuals until the third training condition is met, thus obtaining the second hidden state prediction network. The third training condition could be, for example, that the difference between the third predicted hidden state trajectory and the state trajectory samples of hidden individuals is less than a third difference threshold, or that the number of training iterations reaches a third quantity threshold. It should be noted that the first, second, and third difference thresholds can be different from each other, any two can be the same, or all three can be the same. Similarly, the first, second, and third quantity thresholds can also be different from each other, any two can be the same, or all three can be the same. In other words, the first difference threshold, the second difference threshold, the third difference threshold, and the first quantity threshold, the second quantity threshold, and the third quantity threshold do not affect each other and can be set according to the actual application scenario. For example, the difference between the third predicted hidden state trajectory and the hidden individual's state trajectory sample can be characterized by the MSE (Mean-Square Error) between the third predicted hidden state trajectory and the hidden individual's state trajectory sample.

[0062] In this embodiment, a target structure inference network is used to predict the system structure of the second spliced ​​state trajectory to obtain the second predicted structure features. The second prediction network is then used to predict the hidden trajectory of the second predicted structure features and the state trajectory samples of visible individuals to obtain the third predicted hidden state trajectory. Based on the difference between the third predicted hidden state trajectory and the state trajectory samples of hidden individuals, the parameters of the second prediction network are adjusted until the third training condition is met, thus obtaining the second hidden state prediction network. This network can combine the structure prediction results to predict the hidden state, thereby improving the accuracy of the hidden state prediction and thus improving the accuracy of the second hidden state prediction network.

[0063] In some embodiments, a third predicted hidden state trajectory is obtained by predicting the hidden trajectory of the second predicted structural features and the state trajectory samples of visible individuals through a second prediction network, including:

[0064] Based on the individual combination type, the adjacency matrix corresponding to the second predicted structural feature is divided into blocks to obtain matrix blocks corresponding to multiple combination types. The visible individual matrix blocks within these multiple combination type matrix blocks are then subjected to structural constraints using a multi-head attention mechanism to obtain the constraint penalty value corresponding to the multi-head attention mechanism. The structural constraint strength varies depending on the attention head. The visible individual matrix blocks are used to represent the association between visible individuals. Based on the state trajectory samples of the visible individuals, the initial similarity of the visible individual matrix blocks is determined. Based on the initial similarity and the constraint penalty value, the target similarity corresponding to the multi-head attention mechanism is determined. Based on the target similarity, hidden trajectory prediction is performed to obtain the third predicted hidden state trajectory.

[0065] Here, "individual combination type" refers to the combination type of individual types in the target system. Individual combination types include, for example, visible-visible combinations of visible individuals, visible-hidden combinations of visible and hidden individuals, and hidden-hidden combinations of hidden individuals. The visible individual matrix block is a matrix representing the relationships between visible individuals in the system. The values ​​in the visible individual matrix block are used to represent the relationships between visible individuals.

[0066] For example, the server can divide the adjacency matrix corresponding to the second predicted structural feature into blocks according to the individual combination type. For example, each individual combination type corresponds to one matrix block, resulting in the same number of matrix blocks as the number of combination types.

[0067] In practical applications, for visible individual matrix blocks in matrix blocks corresponding to multiple combination types, structural constraints are applied during the encoding process using a multi-head attention mechanism. This yields the constraint penalty value corresponding to the multi-head attention mechanism. Different attention heads correspond to different constraint strengths, resulting in different constraint penalty values. Based on the constraint penalty value and initial similarity corresponding to the visible individual matrix blocks, the target similarity is determined. Then, during decoding, hidden trajectory prediction is performed based on the target similarity to obtain the third predicted hidden state trajectory.

[0068] In one example, such as Figure 3 As shown, assume that the adjacency matrix corresponding to the obtained second predicted structural feature is Based on the individual combination type, the adjacency matrix can be... Divided into blocks Where N represents the total number of individuals in the target system. This represents the visible individuals matrix, which in turn represents the probability of edges connecting visible individuals. This represents the probability of an edge connecting a visible individual to a hidden individual. This represents the probability of an edge connecting a hidden individual to a visible individual. This represents the probability of an edge connecting two hidden individuals. This represents the second number of hidden individuals. In the self-attention layer of the encoder of the second prediction network, for each attention head i, the unstructured similarity score matrix E is calculated. i (i.e., initial similarity), for example, the dot product between the state trajectory feature vectors of visible individuals can be used as the initial similarity. The constraint penalty value corresponding to each attention head is determined, and the sum of the constraint penalty value and the initial similarity is used as the target similarity for each attention head. Then, the target similarity can be normalized using softmax.

[0069] For example, target similarity .in, Represents the structural guidance coefficients of each attention head; 1 represents the identity matrix with all elements equal to 1; This represents the first number of visible individuals. (Combined) Figure 3 It can be seen that the unconnected node pairs ( The penalty value is close to 0, and is negative, thus suppressing the attention weights of such node pairs after softmax. For node pairs with high connection probabilities, the penalty value is smaller or even 0, making the attention more focused on individuals with structural connections. To balance the model's expressive power and structural prior, for example, a multi-head attention mechanism can be adopted, which includes an attention mechanism with four attention heads, and the corresponding structural guidance coefficients can be [0, -1, -5, -1e9]. That is, the first attention head corresponds to... That is, without using any structural guidance, the original self-attention learning ability is preserved; the second attention head corresponds to And the third attention point corresponding to This corresponds to providing soft structure constraints of varying intensities; the fourth attention head corresponds to... This approximates a hard mask, prohibiting attention between unconnected nodes in the graph. In the decoder of the second prediction network, a pooling-by-multihead-attention (PMA) unit models the interaction between visible and hidden individuals, thus... (Or the symmetrically modified visible-hidden type corresponding matrix block) is mapped to the attention score matrix of PMA, employing the same multi-head, multi-intensity structure penalty strategy as the encoder. This ensures that each hidden individual seed, when aggregating visible individual information, pays more attention to visible individuals with high connection probabilities. When performing SAB (Set Attention Block) self-attention updates on the hidden features, a... The hidden-hidden structural information is constrained by a multi-head structural mask to limit the attention aggregation between hidden individuals. Through this method, the trained second hidden state prediction network utilizes structural estimation as an explicit prior in the attention calculation of each layer, thus forming a closed loop between hidden state prediction and structural inference.

[0070] In this embodiment, the visible individual matrix obtained by dividing the adjacency matrix corresponding to the second predicted structural feature into blocks is subjected to structural constraints by a multi-head attention mechanism. Constraint penalty values ​​of different strengths are applied by multi-head attention. The target similarity is determined based on the initial similarity of the visible individual matrix blocks and the constraint penalty values. The hidden trajectory is predicted based on the target similarity to obtain the third predicted hidden state trajectory. This enables the prediction of the hidden state trajectory to be guided by the structural prediction results in the attention layer, thereby improving the prediction accuracy of the hidden state trajectory.

[0071] In some embodiments, the above method further includes:

[0072] After training the second prediction network for N iterations, every M iterations, the third predicted hidden state trajectory predicted in the previous training round is concatenated with the state trajectory samples of visible individuals to obtain the third concatenated state trajectory; where N and M are both positive integers; the target structure inference network is used to predict the system structure of the third concatenated state trajectory to obtain the third predicted structure features; the second predicted structure features are then replaced with the third predicted structure features.

[0073] It is easy to understand that during the training of the second prediction network, the prediction of the hidden trajectory is guided by the system structure result predicted by the target structure inference network. As the training rounds of the second prediction network increase, the system structure is predicted by the hidden state trajectory predicted by the second prediction network, and the corresponding structure prediction result is obtained. The historical structure prediction results are updated by the structure prediction result, so that the second prediction network can continue to be trained with structure guidance.

[0074] For example, the target structure inference network predicts the system structure of the second spliced ​​state trajectory to obtain the second predicted structure features. The second prediction network then predicts the hidden trajectory using the second predicted structure features and the state trajectory samples of visible individuals to obtain the third predicted hidden state trajectory. Based on the difference between the third predicted hidden state trajectory and the state trajectory samples of hidden individuals, the second prediction network is trained. After N iterations of training, every M iterations, the third predicted hidden state trajectory predicted in the previous training round is spliced ​​with the state trajectory samples of visible individuals to obtain the third spliced ​​state trajectory. The target structure inference network then predicts the system structure of the third spliced ​​state trajectory to obtain the third predicted structure features. The second predicted structure features are replaced with the third predicted structure features, and the second prediction network continues to be trained until the third training condition is met, resulting in the second hidden state prediction network. Here, M and N are both positive integers, for example, M is 45, 48, or 50, and N is 8, 10, or 15, etc.

[0075] In other words, during the first N training iterations of the second prediction network, the second prediction network is trained using the second prediction structural features. After N iterations of training the second prediction network, every M iterations, the hidden state trajectory predicted by the second prediction network is updated with the previous structural prediction result, that is, the second prediction network is trained using the third prediction structural features or the updated third prediction structural features, until the third training condition is met.

[0076] In this embodiment, after training the second prediction network N times, the system structure is re-predicted using the hidden state trajectory predicted by the second prediction network in the current state every M iterations. This yields a new structure prediction result, which replaces the previous one. The second prediction network is then trained with structure guidance. That is, the structure guidance features are updated every M iterations. This improves the accuracy of the structure guidance information during training. In other words, the training of the second prediction network is guided by increasingly accurate structure guidance information, ultimately improving the accuracy of the hidden trajectory prediction of the second hidden state prediction network.

[0077] In an exemplary embodiment, the structure prediction method for hidden individuals can be implemented using a structure prediction system for hidden individuals. A schematic diagram of the implementation process of the structure prediction method for hidden individuals is shown below. Figure 4As shown in the diagram, the structure prediction system for hidden individuals includes a structure-independent hidden state prediction module (HSP-sa), a structure inference module (NRI baseline), and a structure-guided hidden state prediction module (HSP-sg). HSP-sa corresponds to the first hidden state prediction network; NRI corresponds to the target structure inference network; and HSP-sg corresponds to the second hidden state prediction network.

[0078] For example, the structure-independent hidden state prediction module (HSP-sa) is used to estimate the historical trajectory of hidden individuals based solely on the historical trajectories of visible individuals (i.e., the state trajectories of visible individuals) without using any structural information. This module is implemented based on the Set Transformer architecture, and the specific steps are as follows: (1) Input representation. For each training sample (i.e., the training dataset), assume the number of visible individuals is... The number of hidden individuals is Each individual has a state dimension of d at each time step, and the total number of time steps is T. The trajectories of visible individuals are represented as tensors. First, flatten the time and feature dimensions into one dimension, and obtain... Each row corresponds to a state concatenation vector of a visible individual over the entire time series. (2) Encoder. The data is fed into an encoder consisting of two SAB layers. Each SAB layer uses a multi-head self-attention mechanism to model the set of visible individuals, enabling the representation of each individual to incorporate information from other individuals. The encoder then produces a representation matrix. , where D is the internal feature dimension. (3) Decoder. The decoder part is used to generate based on the representation of visible individuals. The trajectory of a hidden individual. First, construct... A learnable "seed vector" Each seed corresponds to a hidden individual. Using a PMA unit, multi-head attention is performed with the seed vector as the query and Z as the key and value to obtain the initial hidden individual features. Subsequently, H is fed into a SAB layer to model the interdependencies between hidden individuals, resulting in updated features. Finally, regarding Each row is mapped to a length of [length missing] through a row-wise feed-forward (rFF) network. The vector is then reshaped into a shape of... Hidden individual trajectory estimation (4) Loss function and training. During the training phase, since the training data is fully observable, the true hidden individual trajectories can be obtained. Because the arrangement of the hidden individuals' numbers is uncertain, a one-to-one matching algorithm (such as the Hungarian algorithm) can be used for prediction. The system finds the matching relationship with the true trajectory that minimizes the total error, and then calculates the mean square error (MSE) loss between the matched trajectory pairs. The parameters of the HSP-sa module are then optimized based on this MSE loss.

[0079] The Structure Inference (NRI) module is used to learn the latent interaction structure of the system and predict future trajectories based on the currently obtained complete individual trajectories (including visible individual trajectories and estimated hidden individual trajectories). This module is implemented using the NRI (Neural Relational Inference) framework. In this example, it accepts hidden trajectory estimates generated by HSP-sa during the pre-training phase. With real visible trajectory The concatenated sequence is used as input, and the output is the initial structure estimate. The results of future trajectory predictions are used to build subsequent structure caches and provide a foundation for the structure guidance module.

[0080] To fully utilize the interactive information provided by the structure inference module, this example designs a structure-guided hidden state prediction module, HSP-sg, based on HSP-sa. This module further refines the hidden individual trajectories given the current structure estimate. The overall structure of this module is similar to HSP-sa, also employing a Set Transformer architecture. However, it explicitly introduces a structure mask in the attention computation of the encoder and decoder to achieve multi-strength structure constraints. Specific implementation details can be found in the corresponding descriptions in the above embodiments and will not be repeated here.

[0081] As illustrated in the example above, there is a chicken-and-egg coupling relationship between hidden state prediction and structure inference. More accurate hidden state prediction is more beneficial for structure inference, and more accurate structure estimation is more conducive to further improving hidden state prediction. Therefore, this example designs an iterative training strategy based on structure caching, including two stages: pre-training and structure-guided iterative refinement.

[0082] (a) Pre-training phase

[0083] S1-1: On a fully observable training dataset, randomly label some individuals in each sample as "visible" and the rest as "hidden". Use the hidden trajectories as the supervision signal for training HSP-sa and train HSP-sa using the MSE loss after matching.

[0084] S1-2: Using the trained HSP-sa, perform hidden trajectory estimation on the visible trajectory of each training sample to obtain... .Will After concatenation, the data is input into the NRI module, and the NRI parameters are optimized according to the standard NRI training method so that it can predict the interaction structure and future trajectory based on the complete trajectory.

[0085] S1-3: After HSP-sa and NRI pre-training are completed, for each sample in the training set, obtain the structure estimate of that sample using HSP-sa+NRI. It is then written into the "structure cache" (A_cache) as the initial structure.

[0086] (II) Structure-guided iterative refinement stage

[0087] S2-1: In the first N epochs (training rounds) of Phase 2, keep the structure cache unchanged and use the structure estimation in the "structure cache" A_cache. HSP-sg is trained using structural guidance signals. Specifically, for each training sample, its visible trajectory is... Corresponding cache structure Input HSP-sg to predict hidden trajectories The HSP-sg parameters are updated using the matched MSE loss.

[0088] S2-2: After every M training epochs, perform a structure cache update: for each sample in the training set, estimate the hidden trajectory using the current HSP-sg output. With visible trajectory The complete trajectory is pieced together and input into the NRI module to obtain a new structure estimate. The structure estimate is then written back to the structure cache unit of the corresponding sample.

[0089] S2-3: Repeat steps S2-1 and S2-2 until the validation set metrics (e.g., hidden state MSE is less than the first difference threshold, structure prediction accuracy is less than the second difference threshold, etc.) no longer improve significantly or the preset number of training rounds is reached.

[0090] Throughout the training process, the structure cache not only maintains a memory of the structure estimate for the training samples, enabling HSP-sg to continuously use increasingly accurate structure information for hidden state prediction in multiple iterations, but also avoids the high cost of rerunning NRI every time the parameters are updated.

[0091] In some real-world scenarios, the number of hidden individuals in the system This is not known beforehand, and there may be differences between different samples. Several sets of HSP-sa and NRI modules can be pre-trained separately, each set corresponding to one candidate number of hidden individuals. ,For example This includes 5 sets of HSP-sa and NRI modules. The training method is the same as that described in the above embodiments. During the inference phase, for a given test sample, for each candidate... The corresponding HSP-sa and NRI modules are used for forward inference to obtain predictions of the future trajectories of visible individuals, and the corresponding prediction errors or likelihood values ​​are calculated. Candidates that perform best in visible trajectory prediction are selected. This serves as an estimate of the number of hidden individuals in the sample, and the corresponding HSP-sg module is used to perform structure-guided hidden trajectory and structure inference. In long-term operation, further statistical analysis of different... The frequency of occurrence of a value can be used to prune or retrain the model family.

[0092] In the above embodiments, by using Set Transformer as the backbone network for hidden state estimation and introducing multi-intensity structural masks, the prior structural information is effectively utilized, making the attention module robust even with incomplete structural estimation, while avoiding the training instability problem that may be caused by a single hard mask. Through structural caching and iterative training mechanisms, the mutual promotion of hidden state prediction and structural inference is achieved. More accurate hidden state estimation provides more reliable input for structural inference, while more accurate structural estimation serves as a priori guidance in the attention layer for further refinement of the hidden state, thus forming a self-reinforcing closed loop. In partially observable scenarios with hidden individuals, it can simultaneously complete the reconstruction of the hidden individual's historical trajectory, complete interactive structural inference, and prediction of all future trajectories. Compared with models that only rely on the trajectories of visible individuals for prediction, it significantly improves the accuracy and precision of structural prediction.

[0093] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps. It is understood that the steps in different embodiments can be freely combined as needed, and all non-contradictory solutions formed by such combinations are within the scope of protection of this application.

[0094] Based on the same inventive concept, this application also provides a structure prediction device for hidden individuals to implement the structure prediction method for hidden individuals described above. The solution provided by this device is similar to the implementation described in the above method. Therefore, the specific limitations of one or more embodiments of the structure prediction device for hidden individuals provided below can be found in the limitations of the structure prediction method for hidden individuals described above, and will not be repeated here.

[0095] In one exemplary embodiment, such as Figure 5 As shown, a structure prediction device 500 for hidden individuals is provided, including: a visible trajectory acquisition module 502, an initial hiding state prediction module 504, a system structure prediction module 506, and a target hiding state prediction module 508, wherein:

[0096] The visible trajectory acquisition module 502 is used to acquire the state trajectory of visible individuals in the target system;

[0097] The initial hidden state prediction module 504 is used to predict the hidden trajectory of the visible individual's state trajectory through the first hidden state prediction network to obtain the initial predicted hidden state trajectory.

[0098] The system structure prediction module 506 is used to concatenate the initial predicted hidden state trajectory with the state trajectory of the visible individuals to obtain the first concatenated state trajectory, and to perform system structure prediction on the first concatenated state trajectory through the target structure inference network to obtain the target predicted structure features.

[0099] The target hidden state prediction module 508 predicts the hidden trajectory of the target by using the second hidden state prediction network to predict the target prediction structural features and the state trajectory of the visible individuals.

[0100] In some embodiments, the apparatus further includes a first prediction network training module, configured to acquire a training dataset before performing hidden trajectory prediction on the state trajectories of visible individuals through a first hidden state prediction network to obtain an initial predicted hidden state trajectory; the training dataset includes state trajectory samples of multiple individuals; a first number of individuals in the training dataset are labeled as visible individuals, and a second number of individuals other than the first number are labeled as hidden individuals; the state trajectory samples of visible individuals are used to perform hidden trajectory prediction on the first prediction network to obtain a first predicted hidden state trajectory; and the parameters of the first prediction network are adjusted according to the difference between the first predicted hidden state trajectory and the state trajectory samples of hidden individuals until a first training condition is met, thereby obtaining the first hidden state prediction network.

[0101] In some embodiments, the apparatus further includes a structure inference network training module, configured to: acquire a training dataset before performing system structure prediction on the first spliced ​​state trajectory through a target structure inference network to obtain target predicted structural features; the training dataset includes state trajectory samples of multiple individuals; a first number of individuals in the training dataset are labeled as visible individuals, and a second number of individuals other than the first number are labeled as hidden individuals; perform hidden trajectory prediction on the state trajectory samples of visible individuals through a first hidden state prediction network to obtain a second predicted hidden state trajectory; splice the second predicted hidden state trajectory with the state trajectory samples of visible individuals to obtain a second spliced ​​state trajectory; perform system structure prediction on the second spliced ​​state trajectory through an initial structure inference network to obtain a first predicted structural feature; and adjust the parameters of the initial structure inference network according to the difference between the first predicted structural feature and the reference structural feature corresponding to the training dataset until a second training condition is met to obtain the target structure inference network.

[0102] In some embodiments, the above-described apparatus further includes a second prediction network training module, configured to, before performing hidden trajectory prediction on the target predicted structural features and the state trajectories of visible individuals through the second hidden state prediction network to obtain the target predicted hidden state trajectory, perform system structure prediction on the second spliced ​​state trajectory through the target structure inference network to obtain the second predicted structural features; perform hidden trajectory prediction on the second predicted structural features and the state trajectory samples of visible individuals through the second prediction network to obtain the third predicted hidden state trajectory; and adjust the parameters of the second prediction network according to the difference between the third predicted hidden state trajectory and the state trajectory samples of hidden individuals until the third training condition is met to obtain the second hidden state prediction network.

[0103] In some embodiments, the second prediction network training module is further configured to: divide the adjacency matrix corresponding to the second prediction structural feature into blocks according to the individual combination type to obtain multiple matrix blocks corresponding to multiple combination types; apply structural constraints of a multi-head attention mechanism to the visible individual matrix blocks in the multiple matrix blocks corresponding to multiple combination types to obtain the constraint penalty value corresponding to the multi-head attention mechanism; the structural constraint strengths corresponding to different attention heads are different; the visible individual matrix blocks are used to represent the association relationship between visible individuals; determine the initial similarity of the visible individual matrix blocks according to the state trajectory samples of the visible individuals; determine the target similarity corresponding to the multi-head attention mechanism according to the initial similarity and the constraint penalty value; and predict the hidden trajectory based on the target similarity to obtain the third predicted hidden state trajectory.

[0104] In some embodiments, the second prediction network training module is further configured to, after training the second prediction network N times, every M iterations, concatenate the third predicted hidden state trajectory predicted in the previous training round of the current training round with the state trajectory samples of visible individuals to obtain the third concatenated state trajectory; N and M are both positive integers; perform system structure prediction on the third concatenated state trajectory through the target structure inference network to obtain the third predicted structure feature; and replace the second predicted structure feature with the third predicted structure feature.

[0105] The modules in the aforementioned structural prediction device for hidden individuals can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0106] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 6 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media. The database stores data related to a structure prediction method for hidden individuals. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it implements a structure prediction method for hidden individuals.

[0107] Those skilled in the art will understand that Figure 6 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0108] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0109] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0110] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0111] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0112] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0113] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0114] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A structural prediction method for hidden individuals, characterized in that, The method includes: Obtain the state trajectories of visible individuals in the target system; The state trajectory of the visible individual is predicted by the first hidden state prediction network to obtain the initial predicted hidden state trajectory. The initial predicted hidden state trajectory is concatenated with the state trajectory of the visible individual to obtain the first concatenated state trajectory. The target structure inference network is then used to predict the system structure of the first concatenated state trajectory to obtain the target predicted structure features. The hidden trajectory prediction of the target is obtained by using a second hidden state prediction network to predict the hidden trajectory of the target prediction structure features and the state trajectory of the visible individuals.

2. The method according to claim 1, characterized in that, Before performing hidden trajectory prediction on the state trajectory of the visible individual using the first hidden state prediction network to obtain the initial predicted hidden state trajectory, the method further includes: A training dataset is obtained, which includes state trajectory samples of multiple individuals; a first number of individuals in the training dataset are marked as visible individuals, and a second number of individuals other than the first number are marked as hidden individuals; The first predicted hidden state trajectory is obtained by performing hidden trajectory prediction on the state trajectory sample of the visible individual through the first prediction network. Based on the difference between the first predicted hidden state trajectory and the state trajectory sample of the hidden individual, the parameters of the first prediction network are adjusted until the first training condition is met, thus obtaining the first hidden state prediction network.

3. The method according to claim 1, characterized in that, Before performing system structure prediction on the first spliced ​​state trajectory through the target structure inference network to obtain the target predicted structure features, the method further includes: A training dataset is obtained, which includes state trajectory samples of multiple individuals; a first number of individuals in the training dataset are marked as visible individuals, and a second number of individuals other than the first number are marked as hidden individuals; The first hidden state prediction network is used to predict the hidden trajectory of the state trajectory sample of the visible individual to obtain the second predicted hidden state trajectory. The second predicted hidden state trajectory is concatenated with the state trajectory sample of the visible individual to obtain the second concatenated state trajectory. The system structure prediction of the second spliced ​​state trajectory is obtained by using the initial structure inference network to predict the first predicted structure feature; Based on the difference between the first predicted structural features and the reference structural features corresponding to the training dataset, the parameters of the initial structural inference network are adjusted until the second training condition is met, thereby obtaining the target structural inference network.

4. The method according to claim 3, characterized in that, Before performing hidden trajectory prediction on the target prediction structural features and the state trajectory of the visible individual using the second hidden state prediction network to obtain the target prediction hidden state trajectory, the method further includes: The target structure inference network is used to predict the system structure of the second spliced ​​state trajectory to obtain the second predicted structure features. The second prediction network is used to predict the hidden trajectory of the second prediction structure features and the state trajectory samples of the visible individuals to obtain the third predicted hidden state trajectory. Based on the difference between the third predicted hidden state trajectory and the state trajectory sample of the hidden individual, the parameters of the second prediction network are adjusted until the third training condition is met, thus obtaining the second hidden state prediction network.

5. The method according to claim 4, characterized in that, The step of predicting hidden trajectories by using a second prediction network on the second predicted structural features and the state trajectory samples of the visible individuals to obtain a third predicted hidden state trajectory includes: Based on the individual combination type, the adjacency matrix corresponding to the second predicted structural feature is divided into blocks to obtain multiple matrix blocks corresponding to combination types. The visible individual matrix blocks in the matrix blocks corresponding to the multiple combination types are subjected to structural constraints of a multi-head attention mechanism to obtain the constraint penalty value corresponding to the multi-head attention mechanism; the structural constraint strength is different for different attention heads; the visible individual matrix blocks are used to represent the association relationship between visible individuals; Based on the state trajectory samples of the visible individuals, determine the initial similarity of the visible individual matrix blocks; The target similarity corresponding to the multi-head attention mechanism is determined based on the initial similarity and the constraint penalty value. Based on the target similarity, a hidden trajectory is predicted to obtain the third predicted hidden state trajectory.

6. The method according to claim 4, characterized in that, The method further includes: After training the second prediction network for N iterations, every M iterations, the third predicted hidden state trajectory predicted in the previous training round of the current training round is concatenated with the state trajectory sample of the visible individual to obtain the third concatenated state trajectory; N and M are both positive integers. The target structure inference network is used to predict the system structure of the third splicing state trajectory to obtain the third predicted structure features. Replace the second predicted structural feature with the third predicted structural feature.

7. A structural prediction device for hidden individuals, characterized in that, The device includes: The visible trajectory acquisition module is used to acquire the state trajectories of visible individuals in the target system; The initial hidden state prediction module is used to predict the hidden trajectory of the visible individual's state trajectory through the first hidden state prediction network to obtain the initial predicted hidden state trajectory. The system structure prediction module is used to concatenate the initial predicted hidden state trajectory with the state trajectory of the visible individual to obtain a first concatenated state trajectory, and to perform system structure prediction on the first concatenated state trajectory through the target structure inference network to obtain the target predicted structure features. The target hidden state prediction module uses a second hidden state prediction network to predict the hidden trajectory of the target prediction structure features and the state trajectory of the visible individuals, thereby obtaining the target prediction hidden state trajectory.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images