Time series data prediction system and method based on node invariant information representation
By decomposing nodes into invariant and time-varying representations in the time-sequence data prediction system, and combining GCRU encoder-decoder and comparison learning, the problem of failure to fully utilize the complex relationship of time-sequence data in the existing technology is solved, and more efficient time series data prediction.
Patent Information
- Application Number
- CN202411469347.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-21
- Publication Date
- 2025-02-18
AI Technical Summary
When processing financial timing data, the prior art fails to fully utilize the complex relationships of timing data, especially the characteristics of non-diffusion data are not taken into account, resulting in the modeled spatiotemporal relationships being rough.
The time series data prediction system based on node invariant information representation is adopted. The time series data is decomposed into node invariant representation and time-varying representation through the data decomposition module. The GCRU encoder-decoder is used for encoding and decoupling, and positive and negative sample pairs are constructed, and the The embedded representation is obtained through comparative learning to predict.
By clearly decoupling static and dynamic features, the model's ability to express complex spatio-temporal graph data is improved, and the stable structural information and time-related behavioral characteristics of nodes can be better captured, thereby improving prediction accuracy.
Smart Images
Figure CN119474800A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of time series data processing in the financial field, and specifically relates to a time series data prediction system and method based on node invariant information representation. Background Art
[0002] In the financial field, there are often complex interactions between multiple time series data, such as stock prices, trading volumes, interest rates, exchange rates and other data that influence each other. It is necessary to comprehensively consider the correlation and dynamic changes of these variables through joint modeling to improve the prediction accuracy of market trends and financial risks. In recent years, with the development of deep learning, many people have used deep learning methods to solve complex time series correlations. Most methods construct an adjacency matrix and make multivariate time series data into a space-time graph. Subsequently, a dynamic diffusion model was proposed based on the GNN method. The model uses two parts, diffusion graph convolution and sequence model, to model spatial and temporal dependencies. Through diffusion graph convolution, the model can effectively capture the spatial dependencies between nodes.
[0003] These methods have good results in the final results, but they have not been fully utilized for complex time series relationships. The main reason is that time series data can be divided into two types: diffusion and non-diffusive. Diffusion data has continuity and propagation properties in time and space, and can be modeled and analyzed by diffusion equations or wave equations. Non-diffusive data focuses more on locality and discreteness, and is usually analyzed and processed based on discrete events or independent measurement points. This was not considered in most previous methods. Only diffusion data was considered but non-diffusive signals were ignored, which ultimately led to the modeling of complex spatiotemporal relationships being very rough. In multivariate time series analysis and prediction, different methods and techniques need to be used to process and model different types of signals. Summary of the invention
[0004] The present invention aims to solve the deficiencies of the prior art and provides the following solutions:
[0005] A time series data prediction system based on node invariant information representation, comprising: a data decomposition module, a data decoupling module, a positive and negative sample pair construction module and a data prediction module;
[0006] The data decomposition module is used for the time series data set of the Neo4j 5.13.0 database, and decomposes the time series data set of each listed entity into a node-invariant representation and a time-varying representation;
[0007] The data decoupling module is used to construct a GCRU encoder-decoder, and encode and decouple the node invariant representation and the time-varying representation to obtain a fixed vector representation and a dynamic node representation;
[0008] The positive-negative sample pair construction module constructs positive sample pairs based on the node invariant representation, the time-varying representation, the fixed vector representation and the dynamic node representation, and constructs negative sample pairs based on the node invariant representation, other node representations, the fixed vector representation and the dynamic node representation;
[0009] The data prediction module performs comparative learning on the node invariant representation and the time-varying representation based on the positive sample pairs and the negative samples to obtain an embedded representation, and inputs the embedded representation into the GCRU encoder-decoder for decoding to obtain a time series data prediction result.
[0010] Preferably, the data decomposition module comprises: a data collection unit and a data decomposition unit;
[0011] The data collection unit is used for the time series data set of the Neo4j 5.13.0 database, and each of the sensors is used as a node, and the time series data set is converted into graph structure data according to the node, wherein the time series data set includes: a Neo4j 5.13.0 database, which contains information such as node attributes (listed entity, type, name), etc. A financial network constructed using 451 real financial entities;
[0012] The data decomposition unit extracts features from the graph structure data based on a graph convolutional network, and decomposes the time series data set into a node-invariant representation and a time-varying representation based on the extracted features.
[0013] Preferably, the data decoupling module comprises: a codec construction unit and a coding decoupling unit;
[0014] The codec construction unit constructs the GCRU encoder-decoder using graph convolution operations and gated recurrent units;
[0015] The encoding structure unit encodes the node invariant representation to generate the fixed vector representation, and encodes the time-varying representation to generate the dynamic node representation.
[0016] Preferably, the positive and negative sample pair construction module includes: a positive sample pair construction unit and a negative sample pair construction unit;
[0017] The positive sample pair construction unit selects the node invariant representation and the time-varying representation of any of the nodes, and constructs the positive sample pair through the fixed vector representation and the dynamic node representation;
[0018] The negative sample pair construction unit selects the node invariant representation of any of the nodes and the node representations of other nodes except the selected node, and constructs the negative sample pair through the fixed vector representation and the dynamic node representation; the node representation includes: the node invariant representation and the time-varying representation.
[0019] Preferably, the data prediction module includes: a comparative learning unit and a prediction unit;
[0020] The contrastive learning unit adopts a triplet first-order loss function, and based on the positive sample pairs and the negative samples, performs contrastive learning on the node invariant representation and the time-varying representation obtained by decomposing the time series data set to obtain an embedded representation;
[0021] The prediction unit is used to input the embedded representation into the GCRU encoder-decoder for decoding to obtain a time series data prediction result.
[0022] The present invention also provides a time series data prediction method based on node invariant information representation, the prediction method is applied to any of the above-mentioned prediction systems, and comprises the following steps:
[0023] S1. Obtain a time series data set of each sensor in the road network, and decompose the time series data set of each sensor into a node-invariant representation and a time-varying representation;
[0024] S2. construct a GCRU encoder-decoder, and perform encoding decoupling on the node invariant representation and the time-varying representation to obtain a fixed vector representation and a dynamic node representation;
[0025] S3. constructing a positive sample pair based on the node invariant representation, the time-varying representation, the fixed vector representation and the dynamic node representation, and constructing a negative sample pair based on the node invariant representation, other node representations, the fixed vector representation and the dynamic node representation;
[0026] S4. Based on the positive sample pairs and the negative samples, the node invariant representation and the time-varying representation are contrastively learned to obtain an embedded representation, and the embedded representation is input into the GCRU encoder-decoder for decoding to obtain a time series data prediction result.
[0027] Preferably, the S1 comprises:
[0028] Obtain a time series data set of the Neo4j 5.13.0 database, and take each listed entity as a node, and convert the time series data set into graph structure data according to the node, wherein the time series data set includes: a Neo4j 5.13.0 database, which contains node attributes (listed entity, type, name) and other information. A financial network constructed using 451 real financial entities;
[0029] The graph structure data is subjected to feature extraction based on a graph convolutional network, and the time series data set is decomposed into a node-invariant representation and a time-varying representation based on the extracted features.
[0030] Preferably, S2 includes:
[0031] The construction unit uses a graph convolution operation and a gated recurrent unit to construct the GCRU encoder-decoder;
[0032] The node invariant representation is encoded to generate the fixed vector representation, and the time-varying representation is encoded to generate the dynamic node representation.
[0033] Preferably, S3 includes:
[0034] Selecting the node invariant representation and the time-varying representation of any of the nodes, and constructing the positive sample pair through the fixed vector representation and the dynamic node representation;
[0035] The node invariant representation of any of the nodes and the node representations of other nodes except the selected node are selected, and the negative sample pair is constructed through the fixed vector representation and the dynamic node representation; the node representation includes: the node invariant representation and the time-varying representation.
[0036] Preferably, S4 includes:
[0037] Using a triplet first-order loss function, based on the positive sample pairs and the negative samples, the node invariant representation and the time-varying representation obtained by decomposing the time series data set are subjected to comparative learning to obtain an embedded representation;
[0038] The embedded representation is input into the GCRU encoder-decoder for decoding to obtain a time series data prediction result.
[0039] Compared with the prior art, the present invention has the following beneficial effects:
[0040] (1) Most existing graph neural networks (GNNs) encode node features as a whole, without distinguishing between static and dynamic features. These models often assume that node features are either fixed or directly process dynamic features, but do not explicitly decouple these two characteristics. The present invention distinguishes the long-term, global characteristics of nodes from the short-term, time-related behaviors by explicitly decoupling fixed vector representations (static features) and time-varying dynamic node representations. This decoupling method improves the model's ability to express complex spatiotemporal graph data, enabling the model to simultaneously capture the stable structural information of nodes in the graph and their behavioral characteristics that evolve over time. Decoupling helps reduce model complexity, reduce interference, and improve the generalization ability of the model.
[0041] (2) In terms of dynamic graph modeling, the existing technology mainly relies on traditional time series models, such as time series GNN, RNN, LSTM, etc. These models have certain limitations in processing time dependencies, especially when processing large-scale graphs and multi-level time dependencies, some important time series information may be lost. The present invention combines the advantages of graph convolution and recurrent neural network (Recurrent Unit), which can effectively capture the complex spatial relationship between nodes while retaining the dynamic evolution of time series. By embedding graph convolution into the loop structure, the model can simultaneously learn the spatial dependency and temporal dependency of the graph, making the model more flexible and effective when processing tasks such as time series prediction and link prediction.
[0042] (3) Many dynamic graph models rely only on graph structure information and cannot capture the subtle dynamic changes of nodes well. For example, snapshot-based dynamic graph models (such as Temporal GCN) usually perform snapshot modeling of node states through discrete time points, but they may ignore the continuous changes of nodes in the time span between snapshots, resulting in insensitivity to fine-grained temporal changes. The present invention uses dynamic node representation (Dynamic Node Representation), and the above technology can capture the subtle differences of continuous changes of nodes over time. This continuously changing encoding enables the model to better capture short-term detailed dynamics when processing complex time-varying graphs, and is particularly suitable for application scenarios with frequent time changes, such as time series prediction, user behavior changes in social networks, etc.
[0043] (4) In many existing models, the spatial features and time series features of the graph structure are often processed separately. The graph structure is first processed by graph convolution, and then the time information is processed by the time series model. This separate processing method may result in the inability of spatial and temporal information to fully interact, thereby limiting the predictive performance of the model. The GCRU architecture of the present invention tightly integrates graph convolution and recursive neural networks, so that spatial dependencies and temporal dependencies can be processed simultaneously. This fusion method can dynamically update the spatial and temporal features of each node in each time step, thereby improving the ability to capture spatiotemporal correlations between nodes.
[0044] (5) In many graph-based contrastive learning or supervised learning tasks, the construction of positive and negative samples depends only on the graph structure or static features between nodes. This method may not accurately characterize the temporal dynamic relationship between nodes. The present invention can more flexibly construct positive and negative sample pairs by utilizing dynamic and static node representations. For example, positive samples can come from certain specific moments in the time evolution sequence of nodes, while negative samples come from the differences after the nodes evolve over time. This method of combining static and dynamic representations makes the construction of positive and negative sample pairs more reasonable and accurate, and can help the model better distinguish between long-term and short-term changes in node behavior in supervised learning tasks. BRIEF DESCRIPTION OF THE DRAWINGS
[0045] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.
[0046] Figure 1 A schematic diagram of the system structure of an embodiment of the present invention;
[0047] Figure 2 This is a schematic diagram of 451 real main nodes, among which, Figure 2 The nodes in contain different financial entities;
[0048] Figure 3 This is a true prediction curve diagram of an embodiment of the present invention. DETAILED DESCRIPTION
[0049] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0050] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.
[0051] Embodiment 1
[0052] In this embodiment, if Figure 1 As shown, a time series data prediction system based on node invariant information representation includes: a data decomposition module, a data decoupling module, a positive and negative sample pair construction module and a data prediction module.
[0053] The data decomposition module is used to obtain the time series data set of the Neo4j 5.13.0 database and decompose the time series data set of each listed company into node invariant representation and time-varying representation.
[0054] The data decomposition module includes: a data collection unit and a data decomposition unit; the data collection unit is used to obtain the time series data set of each sensor in the highway network, and each sensor is used as a node, and the time series data set is converted into graph structure data according to the node, wherein the time series data set includes: a Neo4j 5.13.0 database, which contains information such as node attributes (listed entity, type, name, etc.). A financial network constructed by 451 real financial entities is adopted; the data decomposition unit extracts features from the graph structure data based on the graph convolutional network, and decomposes the time series data set into node invariant representation and time-varying representation based on the extracted features.
[0055] In this embodiment, the data collection unit obtains a time series data set of a Neo4j 5.13.0 database, wherein the time series data set includes: a Neo4j 5.13.0 database, which contains information such as node attributes (listed entity, type, name, etc.). A financial network constructed by 451 real financial entities is adopted, and each listed entity is taken as a node u. The time series data set is converted into a graph structure data G = (V, E) according to the nodes. In the graph structure data G = (V, E), V is a node set, and each node u∈V has a related feature vector x u Afterwards, the data decomposition unit extracts features from the graph structure data based on the graph convolutional network (GCN), where the operation of each layer in the graph convolutional network can be expressed as follows:
[0056]
[0057] Among them, E (l) represents the node feature matrix of the lth layer, t represents the tth time point, Represents the node feature matrix of the l+1th layer; represents the degree matrix, whose diagonal elements D uuWhen H represents the degree of node u, it includes self-connections; (l) represents the hidden state of the lth layer, W (l) represents the trainable weight matrix of the lth layer, and σ represents the activation function ReLU. Each node u is decomposed into two types of representations: (1) The node-invariant representation f(u) whose position remains unchanged. After l layers of GCN, the function f of node u can be defined as It is used to capture the long-term and stable characteristics of the node, independent of the specific location or time; (2) the time-varying representation g(u,t) reflects the state of the node at a specific time t and captures the short-term or dynamic characteristics of the node. This state is an intermediate state or feature calculated by the recurrent neural network model.
[0058] The data decoupling module is used to construct the GCRU encoder-decoder and encode and decouple the node-invariant representation and time-varying representation to obtain a fixed vector representation and a dynamic node representation.
[0059] The data decoupling module includes: a codec construction unit and a coding decoupling unit; the codec construction unit adopts graph convolution operation and gated recursive unit to construct a GCRU encoder-decoder; the coding structure unit encodes the node invariant representation to generate a fixed vector representation, and encodes the time-varying representation to generate a dynamic node representation.
[0060] Time series Some time series data have the characteristics of graph structure data, and also have time series information. In traditional graph convolutional networks, graph convolution operations are mainly used to learn the spatial dependencies between nodes, but cannot directly process time series data. Recursive units (such as LSTM or GRU) can model time series data, but cannot directly capture the spatial dependencies of graph structures. From this, the graph convolution recursive unit (GCRU) that combines the two is derived to simultaneously process spatial and temporal dependencies in graph structure data.
[0061] In this embodiment, the encoder-decoder construction unit uses graph convolution operations (GCN) and gated recurrent units (GRU) to construct a GCRU encoder-decoder as the basic unit of spatiotemporal modeling:
[0062]
[0063] Among them, X represents the graph convolution operation Input, X∈R N×C ; H represents the graph convolution operation The output of H∈R N ×h ; R represents the real number space; N represents the number of nodes in the graph; C represents the initial dimension of each node before the graph convolution operation; h represents the output dimension of each node after the graph convolution; Θ or W krepresents the kernel parameters approximated by Chebyshev polynomials, W k ∈R K×C×h ; K represents the order; k represents the kth order; σ represents the activation function ReLU; Represents an adaptive matrix. The adaptive matrix dynamically reveals the relationship between nodes and captures the changes in time series data. This enables the prediction model to adapt to different time periods and regions, thereby improving the accuracy of prediction. To achieve this, this embodiment projects the node representation E and uses the learnable parameter W 1 , W 2 Get the query matrix E 1 and bond matrix E 2 :
[0064]
[0065] The learnable node representation E contains invariant information. 1 ∈R N×e With E 2 ∈R N×e As the embedded representation of the source node and the target node, the calculated results are normalized to obtain the adaptive adjacency matrix:
[0066]
[0067] Among them, e represents the number of dimensions of the source node and target node embedding representation, represents the adaptive adjacency matrix, represents another adaptive adjacency matrix, T represents the transpose of the matrix, {A} {t} Represents the adaptive adjacency matrix, and α represents the saturation rate of the activation function. The ReLU activation function is used to eliminate weak connections and enhance the robustness of the graph. At the same time, the Softmax function normalizes the non-negative part of the adjacency matrix through random walks to ensure the accurate representation of node relationships. After that, the encoding structure unit encodes the node invariant representation and generates a fixed vector representation. At the same time, the time-varying representation is encoded, and the GCRU unit is used to process the sequence data moment by moment to generate a dynamic node representation.
[0068] The positive and negative sample pair construction module constructs positive sample pairs based on node invariant representation, time-varying representation, fixed vector representation and dynamic node representation, and constructs negative sample pairs based on node invariant representation, other node representations, fixed vector representation and dynamic node representation.
[0069] The positive and negative sample pair construction module includes: a positive sample pair construction unit and a negative sample pair construction unit; the positive sample pair construction unit selects the node invariant representation and time-varying representation of any node, and constructs a positive sample pair through a fixed vector representation and a dynamic node representation; the negative sample pair construction unit selects the node invariant representation of any node, and the node representation of other nodes except the selected node, and constructs a negative sample pair through a fixed vector representation and a dynamic node representation; the node representation includes: a node invariant representation and a time-varying representation.
[0070] Both fixed vector representations and dynamic node representations can be used to construct positive and negative sample pairs. Fixed representations capture static characteristics between nodes, while dynamic representations capture temporal changes. By combining the two, the model can generate more meaningful positive and negative sample pairs, helping the model distinguish between long-term and short-term dependencies between nodes.
[0071] In order to further optimize the effect of signal decomposition, contrastive learning can be used to enhance the distinction between different representations: The positive sample pair construction unit constructs positive sample pairs. For each node u, the representations at different time points t are selected to form positive sample pairs:
[0072] (f(u), g(u, t));
[0073] These sample pairs should be close in the feature space, indicating the similarity of the same node at different times. The negative sample pair construction unit constructs negative sample pairs, and the node invariant representation of node u and the representation of other nodes v (and u≠v) (including node invariant representation and time-varying representation) are selected as negative sample pairs:
[0074]
[0075] These pairs of samples should be far apart in feature space, representing different nodes, even though they may be close in space.
[0076] The data prediction module compares the node invariant representation and time-varying representation based on positive sample pairs and negative samples to obtain an embedded representation, and then inputs the embedded representation into the GCRU encoder-decoder for decoding to obtain the time series data prediction result.
[0077] The data prediction module includes: a contrastive learning unit and a prediction unit; the contrastive learning unit adopts a triplet first-order loss function, and based on positive sample pairs and negative samples, contrasts the node invariant representation and time-varying representation obtained by decomposing the time series data set to obtain an embedded representation; the prediction unit is used to input the embedded representation into the GCRU encoder-decoder for decoding to obtain the time series data prediction result.
[0078] The contrastive learning unit uses a triplet first-order loss function for contrastive learning, as shown in the following formula:
[0079]
[0080] in, represents the triplet first-order self-supervised loss function, S represents the total number of sequences (i.e., samples) in the training dataset, N represents the number of nodes, t represents the tth time point, and i represents the i-th node. Hide the representation Project it into the same feature space as the node representation E and convert the corresponding node representation E (i) As positive samples, we represent other nodes as (j) are considered as negative samples. The non-negative margin δ represents the minimum difference between the distance of the positive sample and the distance of the negative sample, which is necessary to make the loss 0. Larger margins penalize negative samples that are not far enough away from the anchor point to be positive. α controls the lower limit of the distance difference between positive and negative samples. By setting this difference, it is ensured that the distance between the negative sample and the anchor point is always greater than the distance between the positive sample and the anchor point. The mean absolute error (MAE), RMSE, and MAPE are then used as evaluation indicators for the prediction task, and the method is optimized by minimizing the following loss function:
[0081]
[0082] in, represents the predicted regression loss function, ρ represents a time step, represents a certain predicted value, X t+ρ represents a real value; by treating the contrastive learning of node representation as an auxiliary task, we can reveal the intrinsic characteristics of node positions and reduce the error in downstream time series prediction tasks. The present invention combines the predicted regression loss with the self-supervised contrastive loss to obtain the overall loss:
[0083]
[0084] in, Represents the overall loss function, and λ represents a hyperparameter, which is used to control the weight between the two loss functions. Then, the overall loss function is used to optimize the GCRU encoder-decoder of this method using the back propagation algorithm. The prediction unit is used to input the embedded representation into the GCRU encoder-decoder for decoding to obtain the time series data prediction result.
[0085] Embodiment 2
[0086] The Neo4j 5.13.0 database used in this paper contains node attributes (listed entity, type, name) and other information. A financial network constructed by 451 real financial entities is used. Figure 2 This is a schematic diagram of 451 real main nodes, among which, Figure 2 The nodes in contain different financial entities.
[0087] The specific properties of the above two datasets are shown in Table 1:
[0088] Table 1
[0089]
[0090] This algorithm is compared with Graph WaveNet, DGCRN and MegaCRN, including:
[0091] 1) Graph WaveNet This model uses graph convolution to obtain spatial dependencies, and adds a network completion method, that is, adaptively learning a new adjacency matrix based on the data; in terms of obtaining temporal dependencies, it uses dilated convolution, which does not require the loop and recursion of RNNs, and is implemented using a simple convolutional neural network, which has great advantages in both effect and running time.
[0092] 2) DGCRN is a sequence-to-sequence architecture that replaces the linear units in the traditional GRU with dynamic graph convolutions named DGCRM to capture spatiotemporal dependencies. It can also be noted that at each step of the RNN model, the Graph Generator module is used to dynamically generate a semantically relevant adjacency matrix to capture the spatial variability of the data. This method is a curriculum learning method based on RNN, which not only improves the model performance but also improves the model speed.
[0093] 3) MegaCRN A meta-graph learner for space-time graph (STG) learning to explicitly address spatial and temporal heterogeneity. In addition, this method proposes a general meta-graph convolutional recurrent network (MegaCRN) that relies solely on observational data, is robust, and can adapt to data ranging from normal to non-stationary.
[0094] This experiment is based on real data and we get Figure 3 It is a true prediction curve diagram of the time series data prediction method based on node invariant information representation of the present invention.
[0095] Compared with the performance of other algorithms on the dataset of the present invention, Table 2 shows the performance of the present algorithm and the real time series data. It can be seen that the algorithm of the present invention performs well overall and is significantly better than other algorithms.
[0096] Table 2
[0097]
[0098] Embodiment 3
[0099] In this embodiment, a time series data prediction method based on node invariant information representation includes the following steps:
[0100] S1. Obtain the time series data set of the Neo4j 5.13.0 database, and decompose the time series data set of each listed company into a node-invariant representation and a time-varying representation. S1 includes: obtaining the time series data set of each listed company in the Neo4j 5.13.0 database, taking each listed company as a node, and converting the time series data set into graph structure data according to the node, wherein the time series data set includes: Neo4j 5.13.0 database, which contains information such as node attributes (listed entity, type, name). A financial network constructed using 451 real financial entities; extracting features from the graph structure data based on a graph convolutional network, and decomposing the time series data set into a node-invariant representation and a time-varying representation based on the extracted features.
[0101] S2. Construct a GCRU encoder-decoder and encode and decouple the node invariant representation and time-varying representation to obtain a fixed vector representation and a dynamic node representation. S2 includes: a construction unit uses graph convolution operations and gated recurrent units to construct a GCRU encoder-decoder; encodes the node invariant representation to generate a fixed vector representation, and encodes the time-varying representation to generate a dynamic node representation.
[0102] S3. Construct positive sample pairs based on node invariant representation, time-varying representation, fixed vector representation and dynamic node representation, and construct negative sample pairs based on node invariant representation, other node representations, fixed vector representation and dynamic node representation. S3 includes: selecting the node invariant representation and time-varying representation of any node, and constructing positive sample pairs through fixed vector representation and dynamic node representation; selecting the node invariant representation of any node, and the node representation of other nodes except the selected node, and constructing negative sample pairs through fixed vector representation and dynamic node representation; node representation includes: node invariant representation and time-varying representation.
[0103] S4. Based on the positive sample pairs and negative samples, the node invariant representation and the time-varying representation are contrasted and learned to obtain an embedded representation, and the embedded representation is input into the GCRU encoder-decoder for decoding to obtain the time series data prediction result. S4 includes: using a triplet first-order loss function, based on the positive sample pairs and negative samples, the node invariant representation and the time-varying representation obtained by decomposing the time series data set are contrasted and learned to obtain an embedded representation; the embedded representation is input into the GCRU encoder-decoder for decoding to obtain the time series data prediction result.
[0104] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.
Claims
1. A time series data prediction system based on node invariant information representation, characterized in that: include: Data decomposition module, data decoupling module, positive and negative sample pair construction module and data prediction module; The data decomposition module is used to obtain a time series data set from a Neo4j 5.13.0 database, and decompose the time series data set of each listed entity into a node-invariant representation and a time-varying representation; The data decoupling module is used to construct a GCRU encoder-decoder, and encode and decouple the node invariant representation and the time-varying representation to obtain a fixed vector representation and a dynamic node representation; The positive-negative sample pair construction module constructs positive sample pairs based on the node invariant representation, the time-varying representation, the fixed vector representation and the dynamic node representation, and constructs negative sample pairs based on the node invariant representation, other node representations, the fixed vector representation and the dynamic node representation; The data prediction module performs comparative learning on the node invariant representation and the time-varying representation based on the positive sample pairs and the negative samples to obtain an embedded representation, and inputs the embedded representation into the GCRU encoder-decoder for decoding to obtain a time series data prediction result.
2. According to claim 1, a time series data prediction system based on node invariant information representation is characterized in that: The data decomposition module includes: a data collection unit and a data decomposition unit; The data collection unit is used for the time series data set of the Neo4j 5.13.0 database, and each of the sensors is used as a node, and the time series data set is converted into graph structure data according to the node, wherein the time series data set includes: a Neo4j 5.13.0 database, which contains information such as node attributes (listed entity, type, name), etc. A financial network constructed using 451 real financial entities; The data decomposition unit extracts features from the graph structure data based on a graph convolutional network, and decomposes the time series data set into a node-invariant representation and a time-varying representation based on the extracted features.
3. According to claim 2, a time series data prediction system based on node invariant information representation is characterized in that: The data decoupling module includes: a codec construction unit and a coding decoupling unit; The codec construction unit constructs the GCRU encoder-decoder using graph convolution operations and gated recurrent units; The encoding structure unit encodes the node invariant representation to generate the fixed vector representation, and encodes the time-varying representation to generate the dynamic node representation.
4. According to claim 2, a time series data prediction system based on node invariant information representation is characterized in that: The positive and negative sample pair construction module includes: a positive sample pair construction unit and a negative sample pair construction unit; The positive sample pair construction unit selects the node invariant representation and the time-varying representation of any of the nodes, and constructs the positive sample pair through the fixed vector representation and the dynamic node representation; The negative sample pair construction unit selects the node invariant representation of any of the nodes and the node representations of other nodes except the selected node, and constructs the negative sample pair through the fixed vector representation and the dynamic node representation; the node representation includes: the node invariant representation and the time-varying representation.
5. According to claim 2, a time series data prediction system based on node invariant information representation is characterized in that: The data prediction module includes: a comparative learning unit and a prediction unit; The contrastive learning unit adopts a triplet first-order loss function, and based on the positive sample pairs and the negative samples, performs contrastive learning on the node invariant representation and the time-varying representation obtained by decomposing the time series data set to obtain an embedded representation; The prediction unit is used to input the embedded representation into the GCRU encoder-decoder for decoding to obtain a time series data prediction result.
6. A time series data prediction method based on node invariant information representation, the prediction method is applied to the prediction system according to any one of claims 1 to 5, characterized in that: The following steps are involved: S1. Obtain a time series data set from a Neo4j 5.13.0 database, and decompose the time series data set of each listed entity into a node-invariant representation and a time-varying representation; S2. construct a GCRU encoder-decoder, and perform encoding decoupling on the node invariant representation and the time-varying representation to obtain a fixed vector representation and a dynamic node representation; S3. constructing a positive sample pair based on the node invariant representation, the time-varying representation, the fixed vector representation and the dynamic node representation, and constructing a negative sample pair based on the node invariant representation, other node representations, the fixed vector representation and the dynamic node representation; S4. Based on the positive sample pairs and the negative samples, the node invariant representation and the time-varying representation are contrastively learned to obtain an embedded representation, and the embedded representation is input into the GCRU encoder-decoder for decoding to obtain a time series data prediction result.
7. The method for predicting time series data based on node invariant information representation according to claim 6, characterized in that: The S1 includes: Obtain a Neo4j 5.13.0 database, and use each listed entity as a node, and convert the time series data set into graph structure data according to the node, wherein the time series data set includes: a Neo4j5.13.0 database, which contains node attributes (listed entity, type, name) and other information. A financial network constructed using 451 real financial entities; The graph structure data is subjected to feature extraction based on a graph convolutional network, and the time series data set is decomposed into a node-invariant representation and a time-varying representation based on the extracted features.
8. The method for predicting time series data based on node invariant information representation according to claim 7, characterized in that: The S2 includes: The construction unit uses a graph convolution operation and a gated recurrent unit to construct the GCRU encoder-decoder; The node invariant representation is encoded to generate the fixed vector representation, and the time-varying representation is encoded to generate the dynamic node representation.
9. The method for predicting time series data based on node invariant information representation according to claim 7, characterized in that: The S3 includes: Selecting the node invariant representation and the time-varying representation of any of the nodes, and constructing the positive sample pair through the fixed vector representation and the dynamic node representation; The node invariant representation of any of the nodes and the node representations of other nodes except the selected node are selected, and the negative sample pair is constructed through the fixed vector representation and the dynamic node representation; the node representation includes: the node invariant representation and the time-varying representation.
10. The method for predicting time series data based on node invariant information representation according to claim 7, characterized in that: The S4 includes: Using a triplet first-order loss function, based on the positive sample pairs and the negative samples, the node invariant representation and the time-varying representation obtained by decomposing the time series data set are subjected to comparative learning to obtain an embedded representation; The embedded representation is input into the GCRU encoder-decoder for decoding to obtain a time series data prediction result.