Model construction method and device, equipment and storage medium

Abstract

The present disclosure relates to the technical field of intelligent model, and particularly relates to a model construction method and device, equipment and a storage medium. The method comprises: acquiring spatio-temporal data, and acquiring task parameters required for model construction; based on the task parameters, performing standardization processing on the spatio-temporal data to obtain spatio-temporal graph data expressed in a standardized form; based on the spatio-temporal graph data and the task parameters, constructing a search space corresponding to a target spatio-temporal prediction model; based on the search space and the spatio-temporal graph data, performing network architecture search and network parameter optimization corresponding to the target spatio-temporal prediction model to obtain the target spatio-temporal prediction model. The present disclosure is used to solve the defect that the spatio-temporal data cannot be directly applied to the automatic learning in the spatio-temporal graph modeling task in the prior art, and realizes the process of automatically constructing the target spatio-temporal prediction model based on the spatio-temporal data.

Description

Technical Field

[0001] This disclosure relates to the field of intelligent model technology, and in particular to a model building method, apparatus, device and storage medium. Background Technology

[0002] With the development of next-generation information technologies such as the Internet of Things, 5G, and mobile internet, spatiotemporal data, including vehicle trajectory data, public travel data, and air quality monitoring data, are experiencing explosive growth. Spatiotemporal data can be used for prediction, effectively helping city managers provide decision-making support for the refined management of smart cities. Currently, automated spatiotemporal graph modeling methods based on deep learning can meet the modeling needs of many practical projects. However, existing spatiotemporal graph modeling methods cannot directly apply spatiotemporal data to the automated learning of spatiotemporal graph modeling tasks, thus failing to guarantee the efficient and automated implementation of spatiotemporal graph modeling tasks. Summary of the Invention

[0003] This disclosure provides a model building method, apparatus, device, and storage medium to address the shortcomings of existing technologies in the automated learning of spatiotemporal data that cannot be directly applied to spatiotemporal graph modeling tasks, and to realize the process of automatically building a target spatiotemporal prediction model based on spatiotemporal data.

[0004] This disclosure provides a model construction method, comprising: acquiring spatiotemporal data and acquiring task parameters required for model construction; standardizing the spatiotemporal data based on the task parameters to obtain spatiotemporal graph data expressed in a standardized form; constructing a search space corresponding to a target spatiotemporal prediction model based on the spatiotemporal graph data and the task parameters; and performing network architecture search and network parameter optimization corresponding to the target spatiotemporal prediction model based on the search space and the spatiotemporal graph data to obtain the target spatiotemporal prediction model.

[0005] According to a model construction method provided in this disclosure, the spatiotemporal data includes temporal data and spatial data, and the task parameters include temporal granularity and spatial granularity. The step of standardizing the spatiotemporal data based on the task parameters to obtain spatiotemporal graph data expressed in a standardized form includes: processing the temporal data based on the temporal granularity to obtain a time sample set with consistent granularity; processing the spatial data based on the spatial granularity to obtain a spatial sample set with consistent granularity; and constructing the spatiotemporal graph data expressed in a standardized form based on the time sample set and the spatial sample set.

[0006] According to a model construction method provided in this disclosure, the step of constructing the spatiotemporal graph data in a standardized expression based on the time sample set and the spatial sample set includes: constructing a spatial graph by taking each spatial sampling point in the spatial sample set as a node and the distance between every two nodes as an edge; and / or constructing a spatial graph by taking each spatial sampling point in the spatial sample set as a node and the similarity between every two nodes as an edge; constructing a time-varying sequence corresponding to each spatiotemporal graph based on the time sample set and at least one spatial graph; and obtaining at least one set of spatiotemporal graph data based on at least one spatial graph and the time-varying sequence corresponding to each spatial graph.

[0007] According to a model construction method provided in this disclosure, the step of performing network architecture search and network parameter optimization corresponding to the target spatiotemporal prediction model based on the search space and the spatiotemporal graph data to obtain the target spatiotemporal prediction model includes: when at least two sets of the spatiotemporal graph data are obtained, fusing at least two sets of the spatiotemporal graph data to obtain fused graph data; and performing network architecture search and network parameter optimization corresponding to the target spatiotemporal prediction model based on the search space and the fused graph data to obtain the target spatiotemporal prediction model.

[0008] According to a model construction method provided in this disclosure, the step of processing the time data based on the time granularity to obtain a time sample set with consistent granularity includes: processing the time data based on the time granularity to obtain an initial time sample set; when there is missing data at time sampling points in the initial time sample set, obtaining imputation values ​​based on the time sampling points; and supplementing the initial time sample set with the imputation values ​​to obtain the time sample set.

[0009] According to a model construction method provided in this disclosure, obtaining the fill value based on the time sampling point includes: obtaining the fill value based on the temporal proximity of the time sampling point; or, obtaining the fill value based on the spatial proximity of the time sampling point; or, obtaining the fill value based on the similarity of the time sampling point.

[0010] According to a model construction method provided in this disclosure, the step of constructing a search space corresponding to a target spatiotemporal prediction model based on the spatiotemporal graph data and the task parameters includes: extracting spatiotemporal graph features from the spatiotemporal graph data; determining the network search architecture and at least one operator for the target spatiotemporal prediction model based on the spatiotemporal graph features and the task parameters, wherein the network search architecture includes at least one of a single-chain search space, a fully connected graph search space, and a cell-type search space, and the operator includes at least one of a time operator, a spatial operator, an external factor operator, a normalization operator, and a pooling operator; and constructing the search space corresponding to the target spatiotemporal prediction model based on the network search architecture and at least one of the operators.

[0011] This disclosure also provides a model building apparatus, comprising: an acquisition module for acquiring spatiotemporal data and task parameters required for model building; a standardization processing module for standardizing the spatiotemporal data based on the task parameters to obtain spatiotemporal graph data expressed in a standardized form; a space construction module for constructing a search space corresponding to a target spatiotemporal prediction model based on the spatiotemporal graph data and the task parameters; and a model acquisition module for performing network architecture search and network parameter optimization corresponding to the target spatiotemporal prediction model based on the search space and the spatiotemporal graph data to obtain the target spatiotemporal prediction model.

[0012] This disclosure also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the model building method as described above.

[0013] This disclosure also provides a non-transitory computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the model building method as described above.

[0014] The model building method, apparatus, device, and storage medium disclosed herein acquire spatiotemporal data and the task parameters required for model building. Based on these parameters, the spatiotemporal data is standardized to obtain spatiotemporal graph data expressed in a standardized form. Then, based on the spatiotemporal graph data and the task parameters, a search space corresponding to the target spatiotemporal prediction model is constructed. Finally, based on the search space and the spatiotemporal graph data, a network architecture search and network parameter optimization are performed to obtain the target spatiotemporal prediction model. This process closely links spatiotemporal data with the target spatiotemporal prediction model. After standardizing the spatiotemporal data, the automated construction of the target spatiotemporal prediction model is achieved based on the task parameters and the standardized spatiotemporal graph data, enabling efficient and automated implementation of spatiotemporal graph modeling tasks. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in this disclosure or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is one of the flowcharts illustrating the model building method provided in this disclosure;

[0017] Figure 2 This is a schematic diagram illustrating the knowledge transformation principle of the spatiotemporal data provided in this publication;

[0018] Figure 3 This is an example diagram of the candidate network search architecture and pre-defined operators provided in this publication;

[0019] Figure 4 This is a schematic diagram illustrating the principle of the multi-graph fusion model provided in this publication;

[0020] Figure 5 This is the second flowchart illustrating the model building method provided in this publication;

[0021] Figure 6 This is a schematic diagram of the structure of the model building device provided in this disclosure;

[0022] Figure 7 This is a schematic diagram of the structure of the electronic device provided in this disclosure. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the embodiments of this disclosure.

[0024] With the development of next-generation information technologies such as the Internet of Things, 5G, and mobile internet, spatiotemporal data (e.g., vehicle trajectory data, public travel data, and air quality monitoring data) is experiencing explosive growth. This urban spatiotemporal big data can effectively help city managers perceive the city's operational status in real time, promptly grasp the city's evolution process, and accurately predict the city's future, thus providing a basis for decision-making in the refined management of smart cities. For example, by analyzing urban traffic spatiotemporal big data, traffic conditions can be perceived efficiently in real time, urban traffic operation patterns can be identified, and future traffic conditions can be predicted, thereby providing guidance for urban traffic planning and management.

[0025] Based on the spatiotemporal characteristics of urban tasks, urban spatiotemporal tasks can be further divided into spatiotemporal point modeling tasks (such as sales forecasting), spatiotemporal grid modeling tasks (such as grid traffic forecasting), and spatiotemporal graph modeling tasks (such as road network traffic forecasting).

[0026] For spatiotemporal graph modeling tasks, deep learning-based methods currently outperform other approaches. However, deep learning model design typically requires experienced algorithm experts to design specialized models for a wide range of real-world project needs, resulting in extremely high human and time costs. Therefore, efficient and automated modeling is a key challenge for the large-scale deployment of spatiotemporal graph modeling. Automated network architecture search is a method proposed for deep learning model network design, which can automatically design the optimal network structure for different learning tasks.

[0027] Currently, the commonly used methods for automated network structure search are as follows:

[0028] The first method is the Differentiable Architecture Search (DARTS) method, which is based on gradient optimization. This method achieves differentiable architecture search, but it only provides a gradient-based network architecture optimization method without optimizing for the characteristics of spatiotemporal data. Therefore, it is difficult to achieve good results in processing spatiotemporal data and cannot automatically complete the spatiotemporal graph modeling task for spatiotemporal data.

[0029] The second approach is the Efficient Neural Architecture Search for Spatio-Temporal Prediction (AutoST) method. This method is based on convolutional neural networks, but it is suitable for spatio-temporal grid modeling tasks. Spatio-temporal graph modeling tasks, on the other hand, have more complex inter-node connections than spatio-temporal grids, making it unsuitable for direct application.

[0030] The third approach is the Neural Architecture Search for Predictions of Spatio-Temporal Graph (AutoSTG), which is geared towards spatio-temporal graph modeling tasks. However, this approach works directly on spatio-temporal graphs and cannot effectively address the problems of complex graphs and multi-graph information fusion.

[0031] None of the methods mentioned above provide a basic knowledge transformation process for spatiotemporal data, making it impossible to directly apply spatiotemporal data to automated machine learning for spatiotemporal graph modeling tasks.

[0032] Based on the above analysis, this disclosure provides a model construction method that can be automatically implemented through a software algorithm. This software algorithm can be implemented on any device with data processing capabilities, such as a smartphone, laptop, or server. The following section combines... Figures 1-5 This disclosure describes a model building method according to an embodiment.

[0033] In one embodiment, the process steps of the model building method are as follows:

[0034] Step 101: Obtain spatiotemporal data and the task parameters required for model building.

[0035] In this embodiment, spatiotemporal data refers to data collected based on two dimensions: time and space. For example, vehicle trajectory data collected over a period of time, or air quality monitoring data collected over a period of time.

[0036] In this embodiment, the task parameters are parameters needed during model building and can be customized according to user needs, such as the sampling granularity during spatiotemporal data processing. These task parameters can be obtained in advance through a human-computer interaction device, for example, by inputting task parameters through the device's touchscreen.

[0037] Step 102: Based on the task parameters, the spatiotemporal data is standardized to obtain spatiotemporal graph data expressed in a standardized form.

[0038] In this embodiment, after obtaining the spatiotemporal data, the data is standardized to transform the diverse spatiotemporal data into a standardized spatiotemporal graph. This standardization process facilitates the knowledge transformation of the spatiotemporal data, enabling the automated implementation of the target spatiotemporal prediction model and improving its construction efficiency.

[0039] In one embodiment, the spatiotemporal data includes temporal data and spatial data, and the task parameters include temporal granularity and spatial granularity. When standardizing the spatiotemporal data based on the task parameters to obtain spatiotemporal graph data expressed in a standardized form, the temporal data is processed based on the temporal granularity to obtain a time sample set with consistent granularity; the spatial data is processed based on the spatial granularity to obtain a spatial sample set with consistent granularity; and the standardized spatiotemporal graph data is constructed based on the temporal sample set and the spatial sample set.

[0040] In this embodiment, after acquiring the spatiotemporal data, when standardizing the spatiotemporal data, the granularity of the data is processed to convert the spatiotemporal data into a spatiotemporal data sample set (i.e., a time sample set and a spatial sample set) with consistent temporal and spatial granularity.

[0041] Specifically, the task parameters include temporal granularity and spatial granularity. The temporal granularity is the target granularity for sampling temporal data, determined according to needs and actual conditions; the spatial granularity is the target granularity for sampling spatial data, determined according to needs and actual conditions. Temporal data is processed using the temporal granularity, and similarly, spatial data is processed using the spatial granularity to obtain a time sample set and a spatial sample set with consistent granularity.

[0042] In this embodiment, the initial spatiotemporal data may also be sampled data, but the initial spatiotemporal data granularity differs from the target temporal and spatial granularities. Therefore, when processing the temporal data according to the temporal granularity, if the initial temporal data granularity is too large (i.e., fine-grained data is missing), it can be supplemented using methods such as spline interpolation; if the initial temporal granularity is small, it can be aggregated using methods such as downsampling and averaging to obtain a time sample set divided according to the target temporal granularity.

[0043] Similarly, when processing spatial data according to spatial granularity, if the initial spatial data granularity is too large (i.e., fine-grained data is missing), it can be supplemented by methods such as spline interpolation; if the initial spatial granularity is small, it can be aggregated by methods such as downsampling and averaging to obtain a spatial sample set divided according to the target spatial granularity.

[0044] Based on the above, in a specific example, the task of the target spatiotemporal prediction model is to predict the air quality of a given area for the next hour, given various spatiotemporal data such as historical air quality monitoring values, weather, and regional location sets. During standardization, data is aggregated according to an hourly time interval (temporal granularity) and regional locations (spatial granularity). If the spatial granularity of the air quality monitoring values ​​is larger than the given regional locations (i.e., some areas lack monitoring stations, resulting in missing fine-grained data), methods such as spline interpolation can be used to supplement it. If the temporal granularity of the air quality monitoring values ​​is less than one hour, the average value of the one-hour interval can be taken as the hourly granularity monitoring value.

[0045] In one embodiment, when standardizing time data, there may be data gaps, such as in areas without monitoring stations. In this case, when processing time data based on time granularity to obtain a time sample set with consistent granularity, the process is as follows: Process time data based on time granularity to obtain an initial time sample set; when there are missing data points in the time sampling points of the initial time sample set, obtain imputation values ​​based on the time sampling points; supplement the initial time sample set with the imputation values ​​to obtain the final time sample set.

[0046] Similarly, when processing spatial data based on spatial granularity to obtain a spatial sample set with consistent granularity, the process is as follows: Process spatial data based on spatial granularity to obtain an initial spatial sample set; when there are missing data points in the spatial sampling points of the initial spatial sample set, obtain imputation values ​​based on the spatial sampling points; supplement the initial spatial sample set with the imputation values ​​to obtain the spatial sample set.

[0047] In this embodiment, the filling value is obtained by using time sampling points in the initial time sample set and by using spatial sampling points in the initial spatial sample set. While ensuring the validity of the data, the final time sample set is made more complete, which is conducive to improving the construction efficiency of the target spatiotemporal prediction model.

[0048] In one embodiment, the filling value is obtained based on the time sampling point, which can be achieved by: obtaining the filling value based on the temporal proximity of the time sampling points; or, obtaining the filling value based on the spatial proximity of the time sampling points; or, obtaining the filling value based on the similarity of the time sampling points.

[0049] Similarly, the filling value can be obtained based on spatial sampling points in the following ways: based on the spatial proximity of spatial sampling points; or based on the spatial proximity of spatial sampling points; or based on the similarity of spatial sampling points.

[0050] In this embodiment, a filling value is calculated based on a time proximity-based filling method. The formula is as follows:

[0051]

[0052]

[0053]

[0054] in, The fill value is represented by T, which is the window size determined based on proximity; i and j represent different positive integers; y i+j Represents the (i+j)th sampling point (temporal or spatial); c t This is a distance weight based on temporal proximity; for example, a value of 4.

[0055] In this embodiment, a filling value is calculated based on a spatial proximity-based filling method. The formula is as follows:

[0056]

[0057]

[0058] in, The fill value is represented by N, which is the spatial neighborhood set; i and j represent different positive integers; y j Represents the j-th sampling point (time sampling point or spatial sampling point); dist represents the road network distance or Eulerian distance; c s Distance weights are based on spatial proximity.

[0059] In this embodiment, the similarity-based imputation method can calculate the imputation value by weighting the similarity between sampling points (time sampling points or spatial sampling points) (e.g., cosine similarity, Pearson correlation coefficient, etc.).

[0060] In this embodiment, the imputation values ​​of the above-mentioned multiple imputation methods can also be integrated through a simple linear regression model to obtain the final imputation value to fill the missing data.

[0061] In one embodiment, after obtaining a time sample set and a spatial sample set with consistent granularity, spatiotemporal graph data in a standardized expression is constructed based on the time sample set and the spatial sample set. The specific process is as follows: each spatial sampling point in the spatial sample set is used as a node, and the distance between every two nodes is used as an edge to construct a spatial graph; and / or, each spatial sampling point in the spatial sample set is used as a node, and the similarity between every two nodes is used as an edge to construct a spatial graph; based on the time sample set and at least one spatial graph, a time-varying sequence corresponding to each spatiotemporal graph is constructed; based on at least one spatial graph and the time-varying sequence corresponding to each spatial graph, at least one set of spatiotemporal graph data is obtained.

[0062] In this embodiment, spatial sampling points in the spatial data are used as nodes of the spatial graph. Then, the edges of the spatial graph are determined based on the nodes, the required spatial graph is constructed, and a standard spatiotemporal graph representation is output.

[0063] Specifically, the following two methods are provided to define edges to construct a spatial graph:

[0064] The first method constructs a graph based on the distance between nodes (e.g., spatial distance, road network distance, etc.). The edge feature is the distance value between two nodes, or a 0 or 1 value, where 0 indicates that there is no connection between two nodes and 1 indicates that there is a connection between two nodes.

[0065] The second method constructs a graph based on similarity (e.g., time series similarity, spatial attribute similarity, semantic similarity, etc.). The edge feature is the similarity value between two nodes, or a value of 0 or 1, where 0 indicates that there is no connection between the two nodes and 1 indicates that there is a connection between the two nodes.

[0066] Of course, other methods can be used to determine the edges of the spacetime graph depending on the actual situation and needs.

[0067] In addition, you can choose from the various construction methods provided above. Each method can output a corresponding spatiotemporal graph, thus obtaining multiple spatiotemporal graphs corresponding to the same set of spatiotemporal data.

[0068] In this embodiment, after obtaining the spatial map, the corresponding time-varying sequence can be obtained based on the time sample set and the spatial map.

[0069] In one specific embodiment, the spatiotemporal graph data includes spatial graph information. Spatial map information and Time-varying sequences Among them, spatial map information Based on the characteristics of the target task, they can be divided into two categories: single-space diagrams and multi-space diagrams.

[0070] For a single spatial graph, the standardized representation of spatial graph information is as follows:

[0071]

[0072] in, It is of size N v A set of points; Let be the set of edges, where u and v represent distinct nodes, and (u,v) represents the edge between nodes u and v. Static features of a point (D) v (This refers to the dimension of the static features of a point); The static characteristics of the edges (represented by the adjacency matrix, D) e (representing the dimension of the static features of the edge); f uvj ∈F e This represents the j-th attribute value of the edge from u to v. In particular, if the edge does not exist, it is filled with 0.

[0073] For multi-spatial graphs, the standardized representation of spatial graph information is as follows:

[0074]

[0075] in, It is of size N v A set of points; It is the set of edges; Static features of a point (D) v (This refers to the dimension of the static features of a point); The static characteristics of the edges (represented by the adjacency matrix, D) e (N represents the dimension of the static features of the edge); g Indicates the number of spatial diagrams; Let u and v represent the set of edges in the i-th spatial graph, where u and v represent different nodes, and (u,v) represents the edge between nodes u and v. Let $j$ represent the j-th attribute value of the edge from $u$ to $v$ in the $i$-th spatial graph. Specifically, if this edge does not exist in the $i$-th spatial graph, the corresponding position is filled with 0. Also, if the number of attributes in the $i$-th spatial graph is less than $D$... e The insufficient parts are also filled with 0.

[0076] In this embodiment, Time-varying sequences This refers to the readings of each point on a spatial graph as it changes over time, and can generally be represented by tensors. It means that T o D represents the number of time steps. o This represents the dimension of the time series readings.

[0077] In one embodiment, such as Figure 2 As shown, the knowledge transformation process from spatiotemporal data to spatiotemporal graph data is as follows:

[0078] After acquiring spatiotemporal data and task parameters, spatiotemporal processing is performed on the spatiotemporal data. Specifically, the time data and spatial data are granularized to obtain a time sample set and a spatial sample set with consistent granularity. The time sample set and the spatial sample set are collectively referred to as the spatiotemporal data sample set.

[0079] If there are missing data in the spatiotemporal data sample set, missing data can be filled in. Specifically, filling can be based on temporal proximity, spatial proximity, or similarity.

[0080] Next, a spatiotemporal graph is constructed based on the spatiotemporal data sample set. Specifically, the spatiotemporal graph is constructed based on similarity and / or distance.

[0081] Once constructed, a standardized spatiotemporal graph representation is obtained, that is, spatiotemporal graph data expressed in a standardized form.

[0082] Step 103: Based on the spatiotemporal graph data and task parameters, construct the search space corresponding to the target spatiotemporal prediction model.

[0083] In this embodiment, spatiotemporal data is standardized to obtain spatiotemporal graph data. Then, based on the standardized spatiotemporal graph data and task parameters, a search space corresponding to the target spatiotemporal prediction model can be constructed.

[0084] In one embodiment, the specific process for constructing the search space corresponding to the target spatiotemporal prediction model based on spatiotemporal graph data and task parameters is as follows: extracting spatiotemporal graph features from the spatiotemporal graph data; determining the network search architecture and at least one operator for the target spatiotemporal prediction model based on the spatiotemporal graph features and task parameters, wherein the network search architecture includes at least one of a single-chain search space, a fully connected graph search space, and a cell-type search space, and the operator includes at least one of a time operator, a spatial operator, an external factor operator, a normalization operator, and a pooling operator; and constructing the search space corresponding to the target spatiotemporal prediction model based on the network search architecture and at least one operator.

[0085] In this embodiment, the spatiotemporal graph features of the spatiotemporal graph data refer to features such as the number of nodes in the spatiotemporal graph. Task parameters include at least one of the following: user-input prediction step size, indicator type (classification or regression), and number of prediction targets. Then, based on the spatiotemporal graph features and task parameters, the network search architecture and at least one operator of the target spatiotemporal prediction model are determined.

[0086] In this embodiment, multiple candidate network search architectures and multiple operators are pre-set to suit the characteristics of spatiotemporal data. For example... Figure 3 As shown, three candidate network search architectures are pre-defined:

[0087] 1. Single-chain search space: In this network search architecture, graph nodes are connected sequentially through automatically selected operators (Op). This architecture has a smaller search space and is suitable for simpler tasks. Here, Input represents the input, and Output represents the output.

[0088] 2. Fully Connected Graph Search Space: In this network search architecture, graph nodes are fully connected through automatically selected operators. This architecture has the most path connections, and the network depth is unlikely to be too large, making it suitable for relatively simple tasks. Here, Node1, Node2, ... represent different graph nodes.

[0089] 3. Cell-based search space; a combination of singly linked (cells connected sequentially) and fully connected (nodes within a cell are fully connected), integrating the advantages of the two methods mentioned above. Here, Out represents the output of a single cell.

[0090] In this embodiment, the area above the horizontal line represents a variety of built-in operators for handling complex and diverse spatiotemporal correlations, including:

[0091] Temporal operators (T-op) specifically include at least one of the following: 1D convolution (1D Conv) operator, Temporal Convolutional Network (TCN) operator, Long Short-Term Memory (LSTM) operator, gated recurrent neural network (GRU) operator, Multi-Head Attention (MHA) operator, label-based convolution (Meta-Conv) operator, and Recurrent Neural Network (RNN) operator;

[0092] Spatial Operator (S-op) specifically includes at least one of Graph Convolutional Network (GCN) operators, Graph Attention Network (GAT) operators, Graph Multi-Head Attention (G-MHA) operators, and label-based Graph Convolutional Network (Meta-GCN) operators.

[0093] External Operator (E-op) specifically includes fully connected (Dense) operators;

[0094] Normalization operators specifically include at least one of the BatchNorm operator and the LayerNorm operator;

[0095] Pooling operators specifically include one of the convolutional pooling (ConvPool) operator and the max pooling (MaxPool) operator.

[0096] In this embodiment, the network search architecture and at least one operator of the target spatiotemporal prediction model are determined, and the search space of the target spatiotemporal prediction model is automatically constructed.

[0097] Step 104: Based on the search space and spatiotemporal graph data, perform network architecture search and network parameter optimization for the target spatiotemporal prediction model to obtain the target spatiotemporal prediction model.

[0098] In this embodiment, a pre-set selection search algorithm is used to automatically search for and optimize network architecture parameters, ultimately outputting the best model. Specifically, the search algorithm can be any of the following, depending on the actual situation and needs: Differentiable Architecture Search (DARTS), Progressive Neural Architecture Search (PNAS), random search-based optimization methods, reinforcement learning-based network architecture search methods, gradient optimization-based network architecture search methods, or any other search algorithm capable of implementing this method.

[0099] Based on the spatiotemporal graph data obtained in the above embodiments, and the search space of the target spatiotemporal prediction model, a search algorithm is used to perform network architecture search and network parameter optimization to obtain the optimal target spatiotemporal prediction model.

[0100] In one embodiment, if multiple sets of spatiotemporal graph data corresponding to the spatiotemporal data mentioned in the above embodiments are obtained, it is necessary to first use a multi-graph fusion module to fuse the multi-graph knowledge. Specifically, based on the search space and spatiotemporal graph data, the network architecture search and network parameter optimization corresponding to the target spatiotemporal prediction model are performed. When obtaining the target spatiotemporal prediction model, if at least two sets of spatiotemporal graph data are obtained, at least two sets of spatiotemporal graph data are fused to obtain fused graph data. Based on the search space and fused graph data, the network architecture search and network parameter optimization corresponding to the target spatiotemporal prediction model are performed to obtain the target spatiotemporal prediction model.

[0101] In this embodiment, the multi-graph fusion module includes multiple preset graph knowledge learners. Each graph knowledge learner can be pre-configured based on a graph convolutional neural network (GCN), a graph neural network (GNN), or a graph attention network (GAT).

[0102] like Figure 4 As shown, assuming S sets of spatiotemporal graph data are obtained based on spatiotemporal data, S graph knowledge learners are used to learn from each of the S sets of spatiotemporal graph data, obtaining graph knowledge representations corresponding to each of the S sets of spatiotemporal graph data. Finally, a weighted average of the S graph knowledge representations is taken to obtain the fused graph data. Based on the search space and the fused graph data, the network architecture search and network parameter optimization corresponding to the target spatiotemporal prediction model are performed to obtain the target spatiotemporal prediction model.

[0103] Based on the above embodiments, after obtaining the target spatiotemporal prediction model, the spatiotemporal map prediction task is completed using the target spatiotemporal prediction model. Specifically, the spatiotemporal map prediction task refers to a given dynamic sequence of a spatiotemporal map. and corresponding spatial map information Predicting the future spatiotemporal dynamic sequence Among them, D t These are the feature dimensions of the output, T. t These represent the number of time steps output.

[0104] In this embodiment, in multivariate time series prediction problems, the series to be predicted is typically referred to as the target variable, while other relevant observed series are covariates used to assist in the prediction of the target variable. In the above definition, the input dimension D... o It includes target variables and covariates, and outputs feature dimension D. t This represents the dimension of the target variable.

[0105] In an overall embodiment, such as Figure 5 As shown, the implementation process of the model building method is as follows:

[0106] First, knowledge transformation. Specifically, the input spatiotemporal data is standardized into spatiotemporal graph data. This includes pre-defined methods for missing value imputation and spatiotemporal graph construction. Users input spatiotemporal data and task parameters, and select the appropriate processing method (default settings can be preset). The final output is standardized spatiotemporal graph data. Simultaneously, spatiotemporal graph features can be extracted from the spatiotemporal graph data.

[0107] Then, the search space is automatically constructed. Specifically, the selection space for the model structure is configured. Considering the characteristics of spatiotemporal data, this method incorporates components such as time operators, spatial operators, and external factor operators, and pre-configures three candidate network search architectures. The user inputs task parameters such as prediction step size, indicator type (classification or regression), and number of prediction targets, and the search space is automatically constructed.

[0108] Finally, the user selects a search algorithm, the automated network architecture searches and optimizes network parameters, and ultimately outputs the best spatiotemporal prediction model for the target.

[0109] The model building method disclosed herein acquires spatiotemporal data and the task parameters required for model building. Based on these parameters, the spatiotemporal data is standardized to obtain spatiotemporal graph data expressed in a standardized form. Then, based on the spatiotemporal graph data and the task parameters, a search space corresponding to the target spatiotemporal prediction model is constructed. Finally, based on the search space and the spatiotemporal graph data, a network architecture search and network parameter optimization are performed to obtain the target spatiotemporal prediction model. This process closely links spatiotemporal data with the target spatiotemporal prediction model. After standardizing the spatiotemporal data, the automated construction of the target spatiotemporal prediction model is achieved based on the task parameters and the standardized spatiotemporal graph data, enabling efficient and automated implementation of spatiotemporal graph modeling tasks.

[0110] Furthermore, considering the significant differences in spatiotemporal correlations among various spatiotemporal graph prediction tasks, this method accelerates the design and development of spatiotemporal graph models through an automated processing flow from spatiotemporal data to deep spatiotemporal graph models. For different types of spatiotemporal data, this method implements a knowledge transformation model, standardizing the spatiotemporal data into spatiotemporal graph data, providing a foundation for spatiotemporal graph modeling tasks. Addressing the complex and diverse spatiotemporal correlations, this method incorporates multiple operators, including time operators, spatial operators, and external factor operators, and pre-configures three candidate network search architectures, which can be automatically selected based on specific task conditions, comprehensively supporting the needs of spatiotemporal graph network structure search. Simultaneously considering the multi-graph fusion problem, a multi-graph fusion module is designed to fuse multiple sets of spatiotemporal graph data corresponding to the spatiotemporal data, and then construct the target spatiotemporal prediction model through the fused graph data.

[0111] The model building apparatus provided in the embodiments of this disclosure is described below. The model building apparatus described below can be referred to in correspondence with the model building method described above. For example... Figure 6 As shown, the model building apparatus includes:

[0112] The acquisition module 601 is used to acquire spatiotemporal data and the task parameters required for model construction.

[0113] The standardization processing module 602 is used to standardize spatiotemporal data based on task parameters to obtain spatiotemporal graph data expressed in a standardized form.

[0114] The spatial construction module 603 is used to construct the search space corresponding to the target spatiotemporal prediction model based on spatiotemporal graph data and task parameters;

[0115] The model acquisition module 604 is used to search for the network architecture and optimize the network parameters corresponding to the target spatiotemporal prediction model based on the search space and spatiotemporal graph data, so as to obtain the target spatiotemporal prediction model.

[0116] In one embodiment, the standardization processing module 602 is used to process time data based on time granularity to obtain a time sample set with consistent granularity; process spatial data based on spatial granularity to obtain a spatial sample set with consistent granularity; and construct a spatiotemporal graph data expressed in a standardized manner based on the time sample set and the spatial sample set; wherein the spatiotemporal data includes time data and spatial data, and the task parameters include time granularity and spatial granularity.

[0117] In one embodiment, the standardization processing module 602 is used to construct a spatial graph by taking each spatial sampling point in the spatial sample set as a node and the distance between every two nodes as an edge; and / or, to construct a spatial graph by taking each spatial sampling point in the spatial sample set as a node and the similarity between every two nodes as an edge; to construct a time-varying sequence corresponding to each spatiotemporal graph based on the time sample set and at least one spatial graph; and to obtain at least one set of spatiotemporal graph data based on at least one spatial graph and the time-varying sequence corresponding to each spatial graph.

[0118] In one embodiment, the model acquisition module 604 is used to fuse at least two sets of spatiotemporal graph data to obtain fused graph data when at least two sets of spatiotemporal graph data are obtained; and to perform network architecture search and network parameter optimization corresponding to the target spatiotemporal prediction model based on the search space and the fused graph data to obtain the target spatiotemporal prediction model.

[0119] In one embodiment, the standardization processing module 602 is used to process time data based on time granularity to obtain an initial time sample set; when there is missing data at time sampling points in the initial time sample set, it obtains imputation values ​​based on the time sampling points; and supplements the initial time sample set with the imputation values ​​to obtain a time sample set.

[0120] In one embodiment, the standardization processing module 602 is used to obtain the fill value based on the temporal proximity of the time sampling points; or, based on the spatial proximity of the time sampling points; or, based on the similarity of the time sampling points.

[0121] In one embodiment, the space construction module 603 is used to extract spatiotemporal graph features from the spatiotemporal graph data; based on the spatiotemporal graph features and task parameters, determine the network search architecture and at least one operator for the target spatiotemporal prediction model, wherein the network search architecture includes at least one of a single-chain search space, a fully connected graph search space, and a cell-type search space, and the operator includes at least one of a time operator, a spatial operator, an external factor operator, a normalization operator, and a pooling operator; and based on the network search architecture and at least one operator, construct the search space corresponding to the target spatiotemporal prediction model.

[0122] Figure 7An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 7 As shown, the electronic device may include a processor 701, a communication interface 702, a memory 703, and a communication bus 704. The processor 701, communication interface 702, and memory 703 communicate with each other via the communication bus 704. The processor 701 can call logical instructions in the memory 703 to execute a model building method. This method includes: acquiring spatiotemporal data and acquiring task parameters required for model building; standardizing the spatiotemporal data based on the task parameters to obtain spatiotemporal graph data expressed in a standardized form; constructing a search space corresponding to the target spatiotemporal prediction model based on the spatiotemporal graph data and the task parameters; and performing network architecture search and network parameter optimization corresponding to the target spatiotemporal prediction model based on the search space and the spatiotemporal graph data to obtain the target spatiotemporal prediction model.

[0123] Furthermore, the logical instructions in the aforementioned memory 703 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of this disclosure, essentially, or the parts that contribute to the prior art, or parts of the technical solutions, can be embodied in the form of software products. These computer software products are stored in a storage medium and include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this disclosure. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0124] On the other hand, this disclosure also provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions, and when the program instructions are executed by a computer, the computer can execute the model building method provided by the above methods. The method includes: acquiring spatiotemporal data and acquiring task parameters required for model building; standardizing the spatiotemporal data based on the task parameters to obtain spatiotemporal graph data expressed in a standardized form; constructing a search space corresponding to the target spatiotemporal prediction model based on the spatiotemporal graph data and the task parameters; and performing network architecture search and network parameter optimization corresponding to the target spatiotemporal prediction model based on the search space and the spatiotemporal graph data to obtain the target spatiotemporal prediction model.

[0125] In another aspect, this disclosure also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, is implemented to perform the model construction methods provided above. The method includes: acquiring spatiotemporal data and acquiring task parameters required for model construction; standardizing the spatiotemporal data based on the task parameters to obtain spatiotemporal graph data expressed in a standardized form; constructing a search space corresponding to a target spatiotemporal prediction model based on the spatiotemporal graph data and the task parameters; and performing network architecture search and network parameter optimization corresponding to the target spatiotemporal prediction model based on the search space and the spatiotemporal graph data to obtain the target spatiotemporal prediction model.

[0126] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0127] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0128] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this disclosure, and are not intended to limit them. Although this disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this disclosure.

Claims

1. A model construction method, characterized in that, include: Acquire spatiotemporal data, as well as the task parameters required for model building; Based on the task parameters, the spatiotemporal data is standardized to obtain spatiotemporal map data expressed in a standardized form; wherein, the spatiotemporal data is air quality monitoring data collected within a preset time period; Based on the spatiotemporal graph data and the task parameters, a search space corresponding to the target spatiotemporal prediction model is constructed. This construction includes: extracting spatiotemporal graph features from the spatiotemporal graph data; determining the network search architecture and at least one operator for the target spatiotemporal prediction model based on the spatiotemporal graph features and the task parameters, wherein the network search architecture includes at least one of a single-chain search space, a fully connected graph search space, and a cell-type search space, and the operator includes at least one of a time operator, a spatial operator, an external factor operator, a normalization operator, and a pooling operator; and constructing the search space corresponding to the target spatiotemporal prediction model based on the network search architecture and at least one of the operators. Based on the search space and the spatiotemporal graph data, a network architecture search and network parameter optimization are performed for the target spatiotemporal prediction model to obtain the target spatiotemporal prediction model. This process includes: merging at least two sets of spatiotemporal graph data to obtain fused graph data, given at least two sets of spatiotemporal graph data; and performing a network architecture search and network parameter optimization based on the search space and the fused graph data to obtain the target spatiotemporal prediction model. The target spatiotemporal prediction model is a model for predicting air quality in a given area at a preset time.

2. The model construction method according to claim 1, characterized in that, The spatiotemporal data includes time data and spatial data, and the task parameters include time granularity and spatial granularity; The step of standardizing the spatiotemporal data based on the task parameters to obtain spatiotemporal graph data expressed in a standardized form includes: Based on the time granularity, the time data is processed to obtain a time sample set with consistent granularity; Based on the spatial granularity, the spatial data is processed to obtain a spatial sample set with consistent granularity; Based on the time sample set and the spatial sample set, the spatiotemporal graph data is constructed in a standardized expression.

3. The model construction method according to claim 2, characterized in that, The construction of the spatiotemporal graph data in a standardized representation based on the temporal sample set and the spatial sample set includes: A spatial graph is constructed by taking each spatial sampling point in the spatial sample set as a node and the distance between every two nodes as an edge. And / or, using each spatial sampling point in the spatial sample set as a node and the similarity between every two nodes as an edge, a spatial graph is constructed; Based on the set of time samples and at least one of the spatial graphs, a time-varying sequence corresponding to each of the spatiotemporal graphs is constructed; Based on at least one of the spatial maps and the time-varying sequences corresponding to each of the spatial maps, at least one set of spatiotemporal map data is obtained.

4. The model construction method according to claim 2, characterized in that, The step of processing the time data based on the time granularity to obtain a time sample set with consistent granularity includes: Based on the time granularity, the time data is processed to obtain an initial time sample set; When there are missing data points at time sampling points in the initial time sample set, filler values ​​are obtained based on the time sampling points; The initial time sample set is supplemented with the imputation value to obtain the time sample set.

5. The model construction method according to claim 4, characterized in that, The step of obtaining the fill value based on the time sampling point includes: Based on the temporal proximity of the time sampling points, the fill value is obtained; Alternatively, the fill value can be obtained based on the spatial proximity of the time sampling points; Alternatively, the imputation value can be obtained based on the similarity of the time sampling points.

6. A model building apparatus, characterized in that, include: The acquisition module is used to acquire spatiotemporal data, as well as the task parameters required for model building; The standardization processing module is used to standardize the spatiotemporal data based on the task parameters to obtain spatiotemporal map data expressed in a standardized form; wherein, the spatiotemporal data is air quality monitoring data collected within a preset time period; A space construction module is used to construct a search space corresponding to a target spatiotemporal prediction model based on the spatiotemporal graph data and the task parameters. Constructing the search space based on the spatiotemporal graph data and the task parameters includes: extracting spatiotemporal graph features from the spatiotemporal graph data; determining the network search architecture and at least one operator for the target spatiotemporal prediction model based on the spatiotemporal graph features and the task parameters, wherein the network search architecture includes at least one of a single-chain search space, a fully connected graph search space, and a cell-type search space, and the operator includes at least one of a time operator, a spatial operator, an external factor operator, a normalization operator, and a pooling operator; and constructing the search space corresponding to the target spatiotemporal prediction model based on the network search architecture and at least one of the operators. The model acquisition module is used to perform network architecture search and network parameter optimization corresponding to the target spatiotemporal prediction model based on the search space and the spatiotemporal graph data to obtain the target spatiotemporal prediction model. The step of performing network architecture search and network parameter optimization corresponding to the target spatiotemporal prediction model based on the search space and the spatiotemporal graph data to obtain the target spatiotemporal prediction model includes: when at least two sets of the spatiotemporal graph data are obtained, fusing at least two sets of the spatiotemporal graph data to obtain fused graph data; and performing network architecture search and network parameter optimization corresponding to the target spatiotemporal prediction model based on the search space and the fused graph data to obtain the target spatiotemporal prediction model. The target spatiotemporal prediction model is a model for predicting the air quality of a given area at a preset time.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the model building method as described in any one of claims 1 to 5.

8. A non-transitory 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 model building method as described in any one of claims 1 to 5.

Images