CN111754019A - A road segment feature representation learning algorithm based on spatiotemporal graph information maximization model - Google Patents

Abstract

The invention provides a road section feature representation learning algorithm based on a space-time diagram information maximization model, which considers the timeliness of road section states, deeply excavates the time information of road sections, adopts a maximization mutual information mechanism, extracts the mutual influence and interaction relation among the road section information, the time information and the traffic information and utilizes the relation in the learning algorithm based on a neural network. The obtained road section represents the real-time global traffic condition, and learns the upstream and downstream real-time dependency relationship among the road sections, so that the precision of travel time prediction is greatly improved.

Description

A road segment feature representation learning algorithm based on spatiotemporal graph information maximization model

technical field

The invention relates to related fields such as graph neural networks, and more particularly, to a road segment feature representation learning algorithm based on a spatiotemporal graph information maximization model.

Background technique

With the surge in the number of motor vehicles, urban traffic congestion is becoming more and more severe, and a series of problems such as low travel efficiency and waste of resources are introduced. Travel time prediction plays a vital role in applications such as traffic management, route planning, carpooling, and vehicle dispatching. Almost all travel service apps now have this feature, such as Google Maps, Baidu Maps, Didi, etc. With the support of accurate travel time estimates, users can reasonably plan their personal appearance paths and avoid wasting time in congested road sections. At the same time, the city can also reasonably guide the route to effectively alleviate the congestion problem. Therefore, many researchers work on timely and efficient travel time estimation. However, providing accurate estimates remains a challenging task due to the complexity of travel time estimation. One approach that is used in most travel time prediction tasks is segment-based travel time prediction. It can greatly alleviate the risk of data sparsity brought by path-based travel time prediction, and has received a lot of attention. In previous years, there have been many problems in the method of road segment-based travel time prediction. One of the big problems is that the learning of road segment feature representation is not accurate enough, resulting in unsatisfactory prediction accuracy. With the rise of neural networks, Embedding has been applied to road segment feature representation, and has played a great role in extracting road segment static information such as road type, topological relationship between road segments, etc. Among them, there are non-naive network representation learning methods. This method is based on the graph convolutional neural network to learn the network node feature representation with attributes very well. However, the road network is a complex network, and this method still has the following problems on the road network, that is, 1) it will cause the adjacent road segments to be indistinguishable. The goal of the proposed methods is to make the node representation and the adjacent node representation more similar, but in the actual road network, there will be two adjacent road segments only one of which is congested, so they should be more distinguishable; 2) it The interaction of road segments and traffic conditions is not considered. The global traffic conditions will affect each road segment in the road network, and the state of some key road segments also determines the traffic conditions in turn; 3) It does not consider the time-varying of the road segment itself. In different time periods, the status of the road segment may be inconsistent, for example, the road segment is congested in the morning and clear in the afternoon.

SUMMARY OF THE INVENTION

The invention provides a road section feature representation learning algorithm based on a spatiotemporal map information maximization model, which can realize the real-time feature representation of the learned road section.

In order to achieve above-mentioned technical effect, technical scheme of the present invention is as follows:

A road segment feature representation learning algorithm based on a spatiotemporal graph information maximization model, comprising the following steps:

S1: Extract the attributes of the road segment from the road network, generate the initial vector of the road segment, and construct a temporal adjacency matrix based on the trajectory in the historical data;

S2: Use CNN and max-pooling operations on traffic conditions to extract the corresponding traffic state/flow representation;

S3: Input the data of S1, S2 and S3 into the encoder for training to obtain the real-time road segment representation;

S4: Take the obtained road segment representation as the target, and obtain the dynamic representation of the road segment through the fully connected layer.

Further, the specific process of the step S1 is:

S11: Perform data preprocessing, obtain the static attributes of each road section through the road network, and use the three attributes of road section type, number of lanes, and whether it is a one-way road here;

S12: generate the corresponding one-hot vector for the three attributes, and obtain the initial vector R={r ₁ ,r ₂ ,...,r _N } through the fully connected layer after splicing;

S13: Divide the road segment trajectories of the historical data into time segments, and obtain the time adjacency matrix A ^(t) according to the trajectories of different time periods, that is, if in a certain time period, some road segments obtained from the historical trajectories have been driven multiple times and If there is an upstream-downstream relationship, the corresponding road segment has an adjacency relationship, instead of simply determining the adjacency relationship from the topological relationship.

Further, the specific process of the step S2 is:

S21: divide the corresponding city into a grid, and calculate the congestion situation and traffic flow of the corresponding grid;

S22: Input grid data into CNN to obtain the representation of traffic state and traffic flow;

Based on CNN, a representation that can show real-time traffic conditions is learned from grid data; the same method is used to obtain a representation of grid inflow and outflow; the specific calculation formula is as follows:

s ^(t) =CNN(S ^(t) ).

Further, the specific process of the step S3 is:

S31: Using a graph convolutional neural network, obtain the adjacency representation h ^(t) of the road segment from the initial vector of the road segment and the time adjacency matrix;

S32: Through negative sampling, repeat the steps of S31 to obtain the adjacent representation of damaged road sections

S33: Use the readout function to summarize the adjacency representation of the road segment to obtain the global representation g ^(t) of the graph;

S34: Splicing the global representation, traffic state, inflow, and outflow of the graph to obtain a real-time high-order graph induction

S35: Use the obtained adjacency representation, negative sampling adjacency representation, and graph high-order induction to train the model using gradient descent maximization according to the following objective function, and the adjacency representation of the road segment obtained after the training is stabilized is the final road segment representation, and the function formula is as follows :

Further, in fact, based on the Jensen-Shannon divergence between the positive samples and the negative samples, that is, the J-S divergence, to maximize the interaction information between the road segment representation and the global representation, the obtained adjacency representation tends to retain the interaction information represented by the global graph. , find and preserve local-level similarity such as long-distance road segments with similar structural features.

Further, the specific process of the step S4 is:

S41: Considering that the road segment representation has a time period regularity, map the static representation of the road segment and the time of one-hot encoding to a low-dimensional representation through a fully connected layer to obtain a dynamic road segment representation;

S42: Obtain the road segment representation using the stabilized encoder after training;

S43: Use the L loss to minimize the difference between the two, and optimize the parameters of the fully connected layer. After the training is stable, the dynamic representation of the road segment can be obtained based on the fully connected layer;

S44: In fact, according to the periodicity of the state of the road segment itself, and considering the problem of data sparse, the dynamic representation of the road segment is compressed. The main formula is as follows:

H ^(t) = ε(R+A ^(t) )

Compared with the prior art, the beneficial effects of the technical solution of the present invention are:

The present invention considers the temporality of the road section state, deeply excavates the time information of the road section, and adopts the maximizing mutual information mechanism to extract and utilize the mutual influence and interaction relationship among the road section information, time information and traffic information. in learning algorithms based on neural networks. The obtained road segment representation better reflects the real-time global traffic situation, and learns the real-time upstream and downstream dependencies between road segments, which greatly improves the accuracy of travel time prediction.

Description of drawings

Fig. 1 is the flow chart of the method of the present invention;

Figure 2 is a schematic flow chart of the present invention.

Detailed ways

The accompanying drawings are for illustrative purposes only, and should not be construed as limitations on this patent;

In order to better illustrate this embodiment, some parts of the drawings are omitted, enlarged or reduced, which do not represent the size of the actual product;

It will be understood by those skilled in the art that some well-known structures and their descriptions may be omitted from the drawings.

The technical solutions of the present invention will be further described below with reference to the accompanying drawings and embodiments.

As shown in Figure 1, a road segment feature representation learning algorithm based on a spatiotemporal graph information maximization model includes the following steps:

S1: Extract the attributes of the road segment from the road network, generate the initial vector of the road segment, and construct a temporal adjacency matrix based on the trajectory in the historical data;

S2: Use CNN and max-pooling operations on traffic conditions to extract the corresponding traffic state/flow representation;

S3: Input the data of S1, S2 and S3 into the encoder for training to obtain the real-time road segment representation;

S4: Take the obtained road segment representation as the target, and obtain the dynamic representation of the road segment through the fully connected layer.

Further, the specific process of the step S1 is:

S11: Perform data preprocessing, obtain the static attributes of each road section through the road network, and use the three attributes of road section type, number of lanes, and whether it is a one-way road here;

S12: generate the corresponding one-hot vector for the three attributes, and obtain the initial vector R={r ₁ ,r ₂ ,...,r _N } through the fully connected layer after splicing;

S13: Divide the road segment trajectories of the historical data into time segments, and obtain the time adjacency matrix A ^(t) according to the trajectories of different time periods, that is, if in a certain time period, some road segments obtained from the historical trajectories have been driven multiple times and If there is an upstream-downstream relationship, the corresponding road segment has an adjacency relationship, instead of simply determining the adjacency relationship from the topological relationship.

The specific process of step S2 is:

S21: divide the corresponding city into a grid, and calculate the congestion situation and traffic flow of the corresponding grid;

S22: Input grid data into CNN to obtain the representation of traffic state and traffic flow;

Based on CNN, a representation that can show real-time traffic conditions is learned from grid data; the same method is used to obtain a representation of grid inflow and outflow; the specific calculation formula is as follows:

s ^(t) =CNN(S ^(t) ).

The specific process of step S3 is:

S31: Using a graph convolutional neural network, obtain the adjacency representation h ^(t) of the road segment from the initial vector of the road segment and the time adjacency matrix;

S32: Through negative sampling, repeat the steps of S31 to obtain the adjacent representation of the damaged road segment

S33: Use the readout function to summarize the adjacency representation of the road segment to obtain the global representation g ^(t) of the graph;

S34: Splicing the global representation, traffic state, inflow, and outflow of the graph to obtain a real-time high-order graph induction

S35: Use the obtained adjacency representation, negative sampling adjacency representation, and graph high-order induction to train the model using gradient descent maximization according to the following objective function, and the adjacency representation of the road segment obtained after the training is stabilized is the final road segment representation, and the function formula is as follows :

Further, in fact, based on the Jensen-Shannon divergence between the positive samples and the negative samples, that is, the J-S divergence, to maximize the interaction information between the road segment representation and the global representation, the obtained adjacency representation tends to retain the interaction information represented by the global graph. , find and preserve local-level similarity such as long-distance road segments with similar structural features.

The specific process of step S4 is:

S41: Considering that the road segment representation has a time period regularity, map the static representation of the road segment and the time of one-hot encoding to a low-dimensional representation through a fully connected layer to obtain a dynamic road segment representation;

S42: Obtain the road segment representation using the stabilized encoder after training;

S43: Use the L loss to minimize the difference between the two, and optimize the parameters of the fully connected layer. After the training is stable, the dynamic representation of the road segment can be obtained based on the fully connected layer;

S44: In fact, according to the periodicity of the state of the road segment itself, and considering the problem of data sparse, the dynamic representation of the road segment is compressed. The main formula is as follows:

H ^(t) = ε(R+A ^(t) )

As shown in Figure 2, the core purpose of the present invention is the role of road segment representation in traffic prediction. Then we must first study the impact of road segment representation on travel time prediction, and determine the data set. We use the urban travel data set provided by Didi Chuxing's "Gaia" data development plan - the Chengdu and Xi'an data sets, published at https: //gaia.didichuxing.com, and the Harbin dataset, published by DeepGTT. Table 1 shows the number of road segments and trajectories of the three datasets.

Then we need to determine the evaluation criteria for travel time prediction. Here, RMSE and MAE, which are commonly used in this field, are used to represent the prediction effect of the model. That is, when we evaluate the difference between the predicted time of the path and the real time.

According to the judging criteria, we divided the three sets of data sets into training set, validation set and test set, which corresponds to the Didi data set, the training set is the trajectory of the first 17 days, the last 10 days of data is used as the test set, and the rest of the data is used as the verification set For the Harbin data set, the training set is the trajectory of the first 3 days, the data of the last day is used as the test set, and the rest of the data is used as the validation set.

Table 1. Dimensional information and interaction information of the dataset

Before this patent, the commonly used methods of road segment representation learning are all based on non-naive network representation learning algorithms. Although this algorithm learns the static attributes of road segments, it does not consider the time factors of road segments and the traffic factors of road networks. , the impact on the accuracy of travel time prediction is still relatively obvious. All our proposed algorithms use a convolutional neural network-based network representation learning algorithm and maximize mutual information to consider the interactions between road segments, road network globals, and traffic conditions.

To compare with previous methods, we also calculated the RMSE and MAE performance of these methods on three datasets, and the splits of training, validation, and test sets are also consistent with our method.

In addition, in order to test the effect of each part of the model on the model, we conducted ablation experiments and produced the following three model variants: 1) ST-DGI/S: Remove the static information of the road segment, that is, treat the road segment as attributeless 2) ST-DGI/T: ignore the time factor of the road segment; 3) ST-DGI/G: ignore the traffic factor of the road network and only use the inductive representation of the road network map.

Table 2. The performance of various models on the two datasets

It can be seen from the results that our invention has obvious improvement compared with the previous methods, which is largely because the present invention starts from the spatiotemporal attributes of the road segments, and maximizes the extraction of local road segments and global segments by maximizing the mutual information mechanism. The interactive information of the road network is used to learn the relationship between road segments and traffic conditions, as well as the dynamic dependencies between road segments. We can also confirm the importance of three factors to the model effect through ablation experiments. Based on an accurate dynamic segment representation, we were able to improve the accuracy of travel time predictions.

The same or similar reference numbers correspond to the same or similar parts;

The positional relationship described in the accompanying drawings is only for exemplary illustration, and should not be construed as a limitation on this patent;

Obviously, the above-mentioned embodiments of the present invention are only examples for clearly illustrating the present invention, rather than limiting the embodiments of the present invention. For those of ordinary skill in the art, changes or modifications in other different forms can also be made on the basis of the above description. There is no need and cannot be exhaustive of all implementations here. Any modification, equivalent replacement and improvement made within the spirit and principle of the present invention shall be included within the protection scope of the claims of the present invention.

Claims (6)

1. a road segment feature representation learning algorithm based on a spatiotemporal map information maximization model, is characterized in that, comprises the following steps: S1: Extract the attributes of the road segment from the road network, generate the initial vector of the road segment, and construct a temporal adjacency matrix based on the trajectory in the historical data; S2: Use CNN and max-pooling operations on traffic conditions to extract the corresponding traffic state/flow representation; S3: Input the data of S1, S2 and S3 into the encoder for training to obtain the real-time road segment representation; S4: Take the obtained road segment representation as the target, and obtain the dynamic representation of the road segment through the fully connected layer. 2. the road section feature representation learning algorithm based on the spatiotemporal map information maximization model according to claim 1, is characterized in that, the concrete process of described step S1 is: S11: Perform data preprocessing, obtain the static attributes of each road section through the road network, and use the three attributes of road section type, number of lanes, and whether it is a one-way road here; S12: generate the corresponding one-hot vector for the three attributes, and obtain the initial vector R={r ₁ ,r ₂ ,...,r _N } through the fully connected layer after splicing; S13: Divide the road segment trajectories of the historical data into time segments, and obtain the time adjacency matrix A ^(t) according to the trajectories of different time periods, that is, if in a certain time period, some road segments obtained from the historical trajectories have been driven multiple times and If there is an upstream-downstream relationship, the corresponding road segment has an adjacency relationship, instead of simply determining the adjacency relationship from the topological relationship. 3. the road section feature representation learning algorithm based on the spatiotemporal graph information maximization model according to claim 2, is characterized in that, the concrete process of described step S2 is: S21: divide the corresponding city into a grid, and calculate the congestion situation and traffic flow of the corresponding grid; S22: Input grid data into CNN to obtain the representation of traffic state and traffic flow; Based on CNN, a representation that can display real-time traffic conditions is learned from grid data; the same method is used to obtain a representation of grid inflow and outflow; the specific calculation formula is as follows: S ^(t) =CNN(S ^(t) ). 4. the road section feature representation learning algorithm based on the spatiotemporal graph information maximization model according to claim 3, is characterized in that, the concrete process of described step S3 is: S31: Using a graph convolutional neural network, obtain the adjacency representation h ^(t) of the road segment from the initial vector of the road segment and the time adjacency matrix; S32: Through negative sampling, repeat the steps of S31 to obtain the adjacent representation of damaged road sections

S33: Use the readout function to summarize the adjacency representation of the road segment to obtain the global representation g ^(t) of the graph; S34: Splicing the global representation, traffic state, inflow and outflow of the graph to obtain a real-time high-order induction of the graph

S35: Use the obtained adjacency representation, negative sampling adjacency representation, and graph high-order induction to train the model using gradient descent maximization according to the following objective function, and the adjacency representation of the road segment obtained after the training is stabilized is the final road segment representation, and the function formula is as follows :

5. The road segment feature representation learning algorithm based on the spatiotemporal graph information maximization model according to claim 4, characterized in that, in fact, it is based on the Jensen-Shannon divergence between the positive samples and the negative samples, that is, the J-S divergence, which maximizes the The interaction information between the road segment representation and the global representation, then the obtained adjacency representation tends to retain the interaction information represented by the global graph, and finds and preserves local-level similarity such as long-distance road segments with similar structural features. 6. The road segment feature representation learning algorithm based on the spatiotemporal map information maximization model according to claim 5, is characterized in that, the concrete process of described step S4 is: S41: Considering that the road segment representation has a time period regularity, map the static representation of the road segment and the time of one-hot encoding to a low-dimensional representation through a fully connected layer to obtain a dynamic road segment representation; S42: Obtain the road segment representation using the stabilized encoder after training; S43: Use the L loss to minimize the difference between the two, and optimize the parameters of the fully connected layer. After the training is stable, the dynamic representation of the road segment can be obtained based on the fully connected layer; S44: In fact, according to the periodicity of the state of the road segment itself, and considering the problem of data sparse, the dynamic representation of the road segment is compressed. The main formula is as follows:

H ^(t) = ε(R+A ^(t) )

Images