Traffic prediction method based on POI data and text information association learning

Abstract

The invention discloses a traffic prediction method based on correlation learning of POI data and text information, and designs a plug-and-play POI use component for extracting intrinsic distribution features of the POI data according to sparsity and density differences of the POI data and traffic sequence data in a feature dimension and static and dynamic change differences in a time dimension. According to the method, a depth-breadth model thought of a recommendation system is used for reference, feature expression is enhanced by adding a gating mechanism to a prediction output layer, and a POI data and text information association learning network is introduced to measure the similarity between POI category semantic information and traffic sequence features, so that static POI data can reflect space-time characteristics. The method comprises the steps of 1, POI data and text information; step 2, establishing a graph convolution backbone network; step 3, establishing an associated learning network; 4, predicting and outputting; according to the method, the effect of an existing graph convolutional network prediction model is improved on traffic speed data sets of four cities in China.

Description

Technical Field

[0001] The present invention relates to a traffic data prediction method for a traffic system, and proposes a traffic prediction method based on POI data and text information association learning, which supports traffic data prediction tasks in a road network and achieves good prediction results. Background Art

[0002] With the popularity of new modes of travel such as shared bicycles and online ride-hailing, a large amount of original real-time vehicle location data is accumulated with the help of mobile software with positioning functions such as online ride-hailing apps, and through back-end computing and processing, the road network traffic status data is directly or indirectly estimated, such as road flow, average speed of road sections, etc. Then, based on these accumulated traffic status data, the traffic prediction model is trained to predict the future road network traffic status, thereby providing route planning and destination arrival time estimation and other service functions for people's daily travel.

[0003] At present, most of the short-term traffic prediction models are spatiotemporal prediction models based on graph convolutional networks, such as the prediction models OGCRNN, OTSGGCN, DGCN and HGCN proposed in the first four chapters or the existing prediction models STGCN, GWNET and ASTGCN. These spatiotemporal prediction models based on graph convolutional networks have made various improvements to graph convolutional networks to improve feature expression capabilities (such as static road network graph convolution, optimized road network graph convolution, dynamic road network graph convolution), and combined with a variety of time network models to extract spatiotemporal features in traffic sequence data, and finally input this spatiotemporal feature into the prediction output layer to predict the traffic status of the future road network. However, in all model training processes, most graph convolutional network prediction models only use a single type of road network traffic features, such as flow, speed, etc., ignoring the impact of other types of data on the road network on traffic prediction, including cross-modal traffic data such as road type, holiday information, weather sequence data, and point of interest data.

[0004] Holiday information or weather sequence data is aligned with traffic sequence data in time, so these two cross-modal features have only distribution differences with traffic sequence data at the feature dimension level. Multimodal traffic prediction models based on holiday information and weather features are relatively mature. For example, Ge et al. proposed Temporal Graph Convolutional Networks (GTCN) to use holiday information and weather sequence data as additional data and perform feature fusion with the backbone prediction network at the prediction output layer; Zhang et al. proposed Deep Spatial-Temporal Residual Network (ST-ResNet) to input holiday information and weather sequence data into a fully connected network to extract additional features, and fuse multimodal features at the prediction output layer to improve prediction accuracy.

[0005] Point of Interest Data (POI data) mainly counts the number of functional buildings or places within the road network area or around the road section, such as hospitals, schools, scenic spots and other building types. From travel experience and traffic research, it is found that these POI data directly or indirectly affect the operation status of the road network. There are two distribution differences between POI data and road network sequence data: (1) In the feature dimension, the sparsity of POI data and the density of traffic sequence data make the feature distribution of the two very different; (2) In the time dimension, POI data is static and unchanged, while traffic sequence data is dynamic and complex, which makes the two types of data have large temporal heterogeneity. Therefore, there are few studies on the impact of POI static data on traffic prediction in graph convolution prediction models. The classic models include the GTCN model, the Spatiotemporal Multi-Graph Convolution Network (STMGCN) and the Geographical and Semantic spatial-temporal Network (GSNet). STMGCN and GSNet collect POI data on urban grid data and use the similarity of POI features between urban grid areas to construct a road network graph representation to embed additional POI prior knowledge. However, urban grid data destroys the connection mode of road network segments, which conflicts with the short-term traffic prediction task based on road network segments studied in this invention. The GTCN model proposes a similar idea of ​​using POI data: constructing a road network graph representation based on the spatial distance of the road segment and the cosine distance of the road segment POI data, thereby integrating the prior knowledge of POI data into the graph convolution prediction model. However, the above methods using POI data are all non-parametric models, which construct a road network graph representation by utilizing the similarity between POI data. This method cannot effectively compensate for the huge cross-modal data distribution difference between POI static data and traffic sequence dynamic data through parameter learning, resulting in low utilization efficiency and insignificant improvement in prediction results. Summary of the invention

[0006] A traffic prediction method based on POI data and text information association learning is proposed. A plug-and-play POI usage component is designed on the graph convolution prediction model to extract the intrinsic distribution characteristics of POI data. The core idea is borrowed from the wide-deep model of click-through rate data in the recommendation system field. POI data is regarded as a sparse integer two-dimensional tensor data. At the last layer of the overall prediction model, i.e., the prediction output layer, the spatiotemporal characteristics of the graph convolution network output are controlled by the threshold mechanism to enhance the output feature expression ability. In addition, in order to solve the distribution heterogeneity problem between POI static data and traffic sequence dynamic data, a POI data association learning network is designed. It learns the similarity between POI category semantic information and traffic sequence data features through a cross-modal feature distance measurement network, so that the original static POI data becomes dynamic POI data affected by spatiotemporal characteristics, thereby improving the effect of POI data on the graph convolution prediction model. The POI data association learning network is verified to be effective for most existing graph convolution network prediction models on four Chinese city speed datasets.

[0007] The main steps of the present invention include the following:

[0008] Step 1: Define POI data and text information:

[0009] Based on the query function provided by Amap, a POI data set is obtained, and the sampling includes the number of 16 types of POIs, including gas stations, car repairs, shopping malls and supermarkets, ticket offices, logistics services, stadiums, hospitals and special hospitals, hotels, tourist attractions, residential areas, government agencies, schools, subway stations, bus stations, parking lots, and banks. The present invention proposes to obtain the POI data corresponding to the road section based on the POI multiple sampling method, and obtain four offset points by offsetting the original road section end point in four diagonal directions, and count the number of 16 types of POIs in the circular area with a radius of these four offset points to construct POI data:

[0010] POI data ={POI data1 ,POI data2 ,POI data3 ,POI data4}

[0011] Among them, POI data1 ,POI data2 ,POI data3 ,POI data4 Four new sampling points data1, data2, data3, and data4 are obtained by offsetting the original end point of the road section in four diagonal directions, and the POI data are statistically obtained in a circular area centered on these four offset points.

[0012] In addition to obtaining POI data In addition, the present invention also collects POI category information and maps the English words or phrases corresponding to the POI category information into semantic vectors through a semantic encoder. Specifically, the encoder is used to encode the English words or phrases corresponding to the POI category into 2D POI semantic data POI Sema .

[0013] Assuming that the urban road network consists of N road sections, and then T road network data within a period of time are sampled at time intervals, a road network sequence data sample is: Where F is the characteristic length of the data, that is, the number of characteristic types of the road network data. The road network data at time t can be expressed as is the traffic status characteristic data of the i-th road section at time t. The entire road network sequence data can be expressed as Where M is the total number of samples of the road network sequence data. Define the road network graph structure as G = (V, E, A), where V = [1, ..., i, ..., N] is the node set of the road network graph structure, E is the edge set of the non-zero weight of the road network graph structure, and A ∈ R N×N It is the adjacency matrix of the road network graph structure used to represent the connection relationship of the graph, that is, the spatial structure:

[0014]

[0015] The traffic prediction problem can be defined as: Input of historical road network sequence data with known input length T1: It is necessary to predict the road network sequence data of the future T2 time length, and record the predicted value as Output: By constructing the prediction model f θ Learn the features of the input sequence Input to predict the Output sequence, where θ is a learnable parameter of the network. When the model uses the spatial relationship of the adjacency matrix A of the road network graph G, it can be specifically expressed as:

[0016]

[0017] in, is the historical road network sequence data, T1 is the input length; is the predicted future road network sequence data, T2 is the prediction length; f θ is the prediction model, θ is the learnable parameter in the network, and A is the adjacency matrix of the road network graph structure.

[0018] Step 2: Build a graph convolution backbone network:

[0019] Obtain road network sequence data through step 1 After that, the graph convolution network is used to extract spatiotemporal features. The present invention uses a variety of graph convolution network models to build a prediction model. The graph convolution backbone network realizes feature extraction of road network sequence data, that is, mapping the original data into a higher-dimensional feature space to obtain high-dimensional features.

[0020] Here we take the GCN-GRU network as an example.

[0021] The input of the GRU cycle unit is the state feature t-1 at time t-1 and the road network data X at time t. t , and its operation flow is defined as follows:

[0022]

[0023] Where σ is the Sigmoid activation function, ⊙ is the matrix Hadamard product, R t is the reset gate feature, which is used to control H t-1 The degree of influence on the characteristics at time t, is the candidate intermediate state feature at time t, Z t is the update gate feature, which is used to control H t-1 and For the output feature H t The degree of influence, Learn parameters for the fully connected layers and is the bias parameter of the fully connected layer.

[0024] Road network sequence data Input into the GRU cycle unit to get Through the GCN network, we can explore the correlations hidden in the road network data to enhance the spatial feature expression ability of the road network graph.

[0025]

[0026] Among them, A is the adjacency matrix of the road network graph structure, W G is a learnable parameter, and l is the number of GCN-GRU network layers.

[0027] Step 3: Build an associative learning network:

[0028] The association learning network combines the POI obtained in step 1 data 、POI Sema And the spatiotemporal characteristics of the road network obtained in step 2 Further improve the accuracy of traffic prediction. At the same time, in order to further enhance the spatiotemporal characteristics of the road network, a multi-layer linear mapping mechanism is designed. Through this mechanism, information can be extracted from feature spaces of different scales and the information can be aggregated and averaged. This not only retains the important characteristics of the original data, but also enables the model to adaptively select relevant POI data according to the current traffic conditions by introducing the POI semantic importance tensor, thereby improving the expressiveness of the features. The specific process is as follows:

[0029] First, through a series of linear mapping operations, the POI semantic data POI Sema Specifically, the association learning network transforms POI semantic data POI into feature representations of different scales through three linear mapping layers. Sema Mapped to 64-dimensional feature space, 16-dimensional feature space and 8-dimensional feature space respectively, we get

[0030]

[0031] in are the POI semantic features at three different scales, W p1 ,W p2 ,W p3 is a learnable parameter, b p1 ,b p2 ,b p3 is the bias parameter.

[0032] Secondly, the spatiotemporal features of the road network are mapped to 64-dimensional feature space, 16-dimensional feature space and 8-dimensional feature space respectively through three layers of 2D convolution or time convolution:

[0033]

[0034] in The spatiotemporal characteristics of road networks at three different scales.

[0035] Then, the distance features of the two cross-modal features in the same feature space are calculated and concatenated to obtain the multi-scale distance feature dist fea :

[0036]

[0037] dist fea =Concat(dist1,dist2,dist3)

[0038] Among them, dist1, dist2, and dist3 are the absolute values ​​of the differences between cross-modal features of different scales. feaIt is the cross-modal distance feature concatenated from dist1, dist2, and dist3 in the feature dimension.

[0039] Next, dist fea Input into the scaled distance learning network, map its feature length to 1D space, and learn the POI semantic importance tensor POI s-imp :

[0040]

[0041] in is the scaled stitching distance feature.

[0042] Finally, through the Hadamard product of tensors, the POI semantic importance features are used to adaptively select POI data with greater correlation with the current road network features, and the POI semantic information is dynamically embedded into the prediction model, so that the static POI data Become dynamic Improve its feature expression ability. Therefore, based on dynamic The enhanced spatiotemporal characteristics of the road network are:

[0043]

[0044] in To enhance the spatiotemporal characteristics of the road network, For dynamic POI data, POI s-imp is the semantic importance of POI, and M is the total number of samples of road network sequence data.

[0045] Step 4: Predict Output

[0046] Step 3 designs an association learning network based on POI data, which enhances the spatiotemporal expression ability of POI data through the POI semantic importance tensor and obtains enhanced spatiotemporal characteristics of the road network. Will Input into the prediction output layer to obtain the predicted value

[0047]

[0048] Where W O ,b O are learnable parameters.

[0049] Based on the predicted true value With the predicted value Construct the model loss function loss, and update all network parameters through the loss function and back propagation algorithm.

[0050]

[0051] Among them, N is the number of road sections and T2 is the prediction step length. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is a framework diagram of the overall model in the present invention.

[0053] Figure 2 This is a framework diagram of the association learning network in the present invention.

[0054] Figure 3 Traffic prediction results of different prediction models on the Jinan dataset.

[0055] Figure 4 Traffic prediction results of different prediction models on the Xi'an dataset.

[0056] Figure 5 Traffic prediction results of different prediction models on the Shenzhen dataset.

[0057] Figure 6 Traffic prediction results of different prediction models on the Chengdu dataset. DETAILED DESCRIPTION

[0058] Four city traffic speed datasets provided by Didi Chuxing Gaia Project are used, namely Jinan, Xi'an, Shenzhen and Chengdu. These original datasets are the average vehicle speed of the road sections with a sampling frequency of 10 minutes. Jinan has 561 sections (nodes), Xi'an has 792 sections (nodes), Shenzhen has 1174 sections (nodes), and Chengdu has 1676 sections (nodes). 11 comparative methods are used for verification tests, including LSTM, GRU, GCRN, STGCN, OGCRNN, OTSGGCN, GWNET, ASTGCN, DGCN, STSGCN and HGCN. Except for LSTM and GRU, the remaining 9 methods are all traffic prediction models based on graph convolutional networks, and the ASTGCN and DGCN models do not use periodic data modules.

[0059] From the results of the four data sets, it can be seen that the prediction model based on the graph convolutional network has higher prediction accuracy than other non-graph convolutional network prediction models. STGCN and GCRN use static road network graphs for representation, so in most cases, their prediction accuracy is worse than that of other road network graph convolutional prediction models. The optimized road network graph learns more spatial connection relationships than the Hadamard product optimized road network graph, which makes OGCRNN have stronger spatial feature expression ability and higher prediction accuracy than OTSGGCN. OGCRNN and OTSGGCN are significantly better than the two dynamic road network graph convolutional prediction models ASTGCN and STSGCN. The reason is that the dynamic graph spatial expression ability of ASTGCN and STSGCN is not as strong as the optimized road network graph. Since STSGCN uses a temporal and spatial dynamic graph, its spatial feature expression ability is enhanced, so the prediction accuracy of STSGCN is better than ASTGCN.

[0060] DGCN uses a latent Laplacian matrix estimation network based on attention mechanism and LSTM to learn dynamic road network graphs, making its feature expression ability and prediction accuracy stronger than OGCRNN, OTSGGCN, ASTGCN and STSGCN. Since the GWNET model uses a deep graph convolutional network layer, its prediction results are better than the DGCN model. HGCN takes into account the influence of the spatiotemporal characteristics of the road network area on the basis of optimizing the road network graph convolution, and thus achieves the best prediction accuracy on all data sets.

Claims

1. A traffic prediction method based on POI data and text information association learning, characterized in that: The implementation steps of this method are as follows: Step 1: POI data and text information: Based on the query function provided by the map, the POI dataset is obtained, including the number of 16 types of POIs, including gas stations, car repairs, shopping malls and supermarkets, ticket offices, logistics services, stadiums, hospitals and special hospitals, hotels, tourist attractions, residential areas, government agencies, schools, subway stations, bus stations, parking lots, and banks; a POI multi-sampling method is proposed to obtain the POI data corresponding to the road section: THEN data ={THEN data1 ,THEN data2 ,THEN data3 ,THEN data4 } Among them, POI data1 ,POI data2 ,POI data3 ,POI data4 Four new sampling points data1, data2, data3, and data4 are obtained by offsetting the original end point of the road section in four diagonal directions, and the POI data are obtained by counting the circular area centered on these four offset points; Step 2: Build a graph convolution backbone network: Obtain road network sequence data through step 1 After that, we use graph convolutional networks to extract spatiotemporal features, and use a variety of graph convolutional network models to build prediction models. Through graph convolutional networks, we can mine the correlation hidden in the road network data to enhance the spatial feature expression ability of the road network graph. First, we transform the road network data into Input to the linear mapping layer and map it to a high-dimensional space; then, combine it with the graph convolutional network prediction model to learn the spatiotemporal characteristics of the road network data The graph convolution backbone network realizes feature extraction of road network sequence data, that is, mapping the original data into a higher-dimensional feature space; Step 3: Build an associative learning network: The association learning network combines the POI obtained in step 1 data 、POI Sema And the spatiotemporal characteristics of the road network obtained in step 2 Further improve the accuracy of traffic forecasts; at the same time, in order to further enhance the spatiotemporal characteristics of the road network, a multi-layer linear mapping mechanism is designed; through this mechanism, information is extracted from feature spaces of different scales and the information is aggregated to calculate the average value; Step 4: Predict output; The result obtained in step 3 Input into the prediction output layer to obtain the predicted value Where W O ,b O are learnable parameters; Based on the predicted true value With the predicted value Construct the model loss function loss, and update all network parameters through the loss function and back propagation algorithm; Among them, N is the number of road sections and T2 is the prediction step length.

2. The traffic prediction method based on POI data and text information association learning according to claim 1 is characterized in that: Step 1: Get POI data In addition, POI category information is collected, and the English words or phrases corresponding to the POI category information are mapped into semantic vectors through a semantic encoder; the English words or phrases corresponding to the POI category are encoded into 2D POI semantic data POI using the encoder Sema ; The urban road network consists of N road sections, and then T road network data are sampled within a period of time according to the time interval. Then a road network sequence data sample is: Where F is the characteristic length of the data, that is, the number of characteristic types of the road network data. The road network data at time t can be expressed as is the traffic status characteristic data of the i-th road section at the t-th time; the entire road network sequence data is expressed as Where M is the total number of samples of the road network sequence data; the road network graph structure is defined as G = (V, E, A), where V = [1, ..., i, ..., N] is the node set of the road network graph structure, E is the edge set of the non-zero weight of the road network graph structure, and A ∈ R N×N It is the adjacency matrix of the road network graph structure, which is used to represent the connection relationship of the graph, that is, the spatial structure; The traffic prediction problem is specifically expressed as: in, is the historical road network sequence data, T1 is the input length; is the predicted future road network sequence data, T2 is the prediction length; f θ is the prediction model, θ is the learnable parameter in the network, and A is the adjacency matrix of the road network graph structure.

3. The traffic prediction method based on POI data and text information association learning according to claim 1 is characterized in that: In step 3, firstly, a series of linear mapping operations are performed to transform the POI semantic data POI Sema Converted into feature representations of different scales; specifically, the association learning network transforms POI semantic data POI through three layers of linear mapping layers Sema Mapped to 64-dimensional feature space, 16-dimensional feature space and 8-dimensional feature space respectively, we get The spatiotemporal features of the road network are transformed through three layers of 2D convolution or time convolution. Mapped to 64-dimensional feature space, 16-dimensional feature space and 8-dimensional feature space respectively, we get Calculate the distance features of two cross-modal features in the same feature space and concatenate them to obtain the multi-scale distance feature dist fea : dist fea =Concat(dist1,dist2,dist3) Among them, dist1, dist2, and dist3 are the absolute values ​​of the differences between cross-modal features of different scales. fea It is the cross-modal distance feature concatenated by dist1, dist2, and dist3 in the feature dimension, POI Sema is the POI semantic data, It is the spatiotemporal characteristics of the road network; Dist fea Input into the scaled distance learning network to map its feature length to a 1-dimensional space, thereby learning the POI semantic importance tensor POI s-imp ; Then, through the Hadamard product of tensors, the POI semantic importance features are used to adaptively select POI data with greater correlation with the current road network features, and the POI semantic information is dynamically embedded into the prediction model, so that the static POI data Become dynamic Improve its feature expression ability; Therefore, based on dynamic The enhanced spatiotemporal characteristics of the road network are: in To enhance the spatiotemporal characteristics of the road network, For dynamic POI data, POI s-imp is the semantic importance of POI, and M is the total number of samples of road network sequence data.

Images