CN118193628A - Semantic enhancement graph contrast learning method and system for road network representation - Google Patents

Abstract

The present invention relates to a semantically enhanced graph contrast learning method and system for road network representation, and belongs to the technical field of traffic data analysis and machine learning. The present invention provides a semantically enhanced graph contrast learning framework, which effectively combines road attributes and visual information through multimodal feature embedding, and deeply mines and utilizes the characteristics of road network data. A new type of road network data representation is constructed, which not only has high-dimensional rich information, but also can be directly used as the input of a machine learning model. This design makes various road network-based downstream tasks, such as road label classification, speed prediction, etc., more convenient to implement. This scheme integrates the idea of graph contrast learning, and by simulating the topological changes and data missing of the road network, it improves the robustness and generalization ability of the road network data representation, effectively solves the limitations of traditional methods in complex road network analysis, and improves the efficiency of road network data processing.

Description

A semantically enhanced graph comparative learning method and system for road network representation

Technical Field

The present invention belongs to the technical field of traffic data analysis and machine learning, and relates to a semantically enhanced graph comparative learning method and system for road network representation.

Background technique

With the acceleration of urbanization, urban traffic networks have become increasingly complex and changeable, and traditional road network analysis methods are facing more and more challenges. Traditional road network analysis methods often rely on simplified mathematical models and limited data sources, making it difficult to capture and explain the complex dynamics and diversity in road networks. For example, in urban planning and traffic management, a deep understanding of key information such as traffic flow, congestion patterns, and accident risks is crucial, but existing technologies and methods are often unable to effectively process and analyze large-scale, high-dimensional road network data. In addition, with the development of technology, the types and magnitudes of road network data have changed significantly. From GPS trajectories of vehicles to real-time traffic condition monitoring, the richness and diversity of data are unprecedented. However, this also brings new challenges to data processing and analysis. Existing analysis tools and methods often seem to be unable to cope with these big data, especially in situations where real-time monitoring of traffic flows, prediction of traffic trends, and response to traffic emergencies are required. For example, in terms of real-time traffic status monitoring, traditional data analysis methods cannot immediately reflect changes in traffic flow; in terms of road condition prediction, due to the lack of in-depth analysis of big data, it is often impossible to accurately predict future traffic conditions; in terms of traffic accident response, traditional methods also have obvious deficiencies in the timeliness and accuracy of data processing. Therefore, in view of the complex and changeable characteristics of modern urban transportation networks, it is urgent to develop new analysis methods and technologies to better understand and manage these complex transportation systems. These new methods and technologies should be able to effectively cope with large-scale, high-dimensional traffic data to achieve in-depth understanding and precise management of transportation networks. This includes not only efficient processing of existing data, but also effective simulation and prediction of traffic network topology changes and data missing situations. Through this technological advancement, urban planning and traffic management can be supported more comprehensively and accurately, thereby improving the efficiency and safety of urban transportation systems.

Summary of the invention

In view of this, the purpose of the present invention is to provide a semantically enhanced graph contrast learning method and system for road network representation, which provides a framework called SE-GCL (Semantic Enhanced Graph Contrastive Learning), which effectively combines road attributes and visual information through multimodal feature embedding, deeply explores and fully utilizes the characteristics of road network data. It constructs a new type of road network data representation, which not only has high-dimensional rich information, but also can be directly used as input for machine learning models. This design makes various road network-based downstream tasks, such as road label classification, speed prediction, and travel time prediction, more convenient to implement. SE-GCL integrates the idea of graph contrast learning, and improves the robustness and generalization ability of road network data representation by simulating the topological changes and data loss of the road network, effectively solving the limitations of traditional methods in complex road network analysis.

In order to achieve the above object, the present invention provides the following technical solutions:

A semantically enhanced graph contrastive learning method for road network representation is proposed. The method constructs a semantically enhanced graph contrastive learning framework for road network representation, which specifically includes the following stages:

Phase 1: Build a multimodal feature embedding module, which uses the attribute data of the road segment and the street view image data to obtain the attribute embedding (embedding, i.e., low-dimensional representation vector) and visual feature embedding of the road segment respectively, and concatenate them together as the initial embedding of the road segment;

The second stage: construct a graph augmentation module. Based on modeling the road network as a graph, design two different augmentation strategies to impose perturbations on the topology and node features of the road network (graph), thereby generating two different augmented graphs (graph variants).

Phase 3: Build a graph representation learning module. Based on the two graph variants generated by the graph augmentation module, use a graph neural network model to extract the node level, i.e., the representation of the road segment.

Phase 4: Build a contrast optimization module based on semantic enhancement, use the road network characteristics and the mobility semantics in the historical trajectory to sample sample pairs, and design a dual-view contrast loss function to guide the contrast optimization process.

Furthermore, in the first stage, it is mainly divided into two parts: attribute feature embedding and visual feature embedding;

The attribute feature embedding part includes: for any road segment v _i , comprehensively analyze its road segment identifier ID, road segment length LEN, road segment midpoint coordinates (LON, LAT) and other attributes, and the symbols corresponding to these attributes are v _i .id, v _i .len, v _i .lon, and v _i .lat respectively; considering that the changes in longitude and latitude coordinates within a single road segment are very limited, it is only necessary to select the longitude and latitude of the geometric midpoint of the road segment to represent the location information of the road segment. For the real number type attributes v _i .lon and v _i .lat, they are converted into discretized forms through binning operations, and then a separate linear layer is used to map each attribute value to a corresponding feature vector. Subsequently, the four feature vectors are concatenated together to form the attribute embedding of the road segment;

The visual feature embedding part includes: using the pre-trained Swin Transformer as the visual feature encoder, taking the panoramic image data collected at the midpoint of the road section as input, and outputting its visual embedding Finally, the two different feature types are concatenated together to obtain the initial embedding of the road segment.

Furthermore, in the second stage, a graph augmentation module is constructed to augment the road network using two strategies: mobility-based edge removal and modality-based feature masking. The mobility-based edge removal strategy targets the topological structure of the road network and determines the edge removal probability based on the transition probability between road segments in the historical trajectory data. If two road segments often appear together in historical trajectories, the connection between them is considered strong and important, so the probability of removing such a connection is very low. The modality-based feature masking strategy is a perturbation strategy for node features, which randomly selects one of the attributes or image features for masking, so that the contrastive learning model can adapt to the lack of road segment data, thereby enhancing the robustness of the model. In each training process, these two different augmentation strategies are used simultaneously to generate two graph views for comparison.

Furthermore, in the third stage, a graph representation learning module is constructed, and the two graph views generated by the graph augmentation module are used as input to learn the representations of the two graph views respectively; specifically, the module includes: using a graph encoder to learn and extract high-level features from the graph views, thereby generating a complex representation of each road segment, and finally, mapping the road segment representation generated by the graph encoder to a low-dimensional latent space through a nonlinear projection head.

Furthermore, in the fourth stage, a contrast optimization module based on semantic enhancement is constructed, which includes a semantic enhancement sampling strategy and can construct positive and negative samples by using the road network characteristics and the mobility semantics in the historical trajectory. Specifically, given a target node (i.e., anchor node) of a certain view, its neighbor nodes within the h-hop are regarded as positive samples within the view, and the mirror nodes of another view are regarded as positive samples between views. For nodes beyond the h-hop range, those nodes that co-occur with the anchor node in the historical trajectory are further screened out, and the remaining nodes are selected as true negative samples. Then, a contrast loss function designed based on the InfoNCE objective is adopted:

To calculate the contrast loss of all nodes on the two graph views, and use this to get the final dual-view contrast loss function

Among them, _v'i represents the mirror node of _vi in another view; finally, this loss function is used to train the network to optimize the road network embedding representation.

The present invention also provides a semantically enhanced graph comparative learning system for road network representation.

The beneficial effects of the present invention are:

The present invention provides a framework called SE-GCL (Semantic Enhanced Graph Contrastive Learning), which effectively combines road attributes and visual information through multimodal feature embedding, deeply explores and fully utilizes the characteristics of road network data. It constructs a new type of road network data representation, which not only has high-dimensional rich information, but also can be directly used as input for machine learning models. This design makes various road network-based downstream tasks, such as road label classification, speed prediction, and travel time prediction, more convenient to implement. SE-GCL incorporates the idea of graph contrastive learning, and improves the robustness and generalization ability of road network data representation by simulating topological changes and data loss in road networks, effectively solving the limitations of traditional methods in complex road network analysis, and improving the efficiency of road network data processing and the understanding of the complex dynamics of urban transportation networks.

Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description.

BRIEF DESCRIPTION OF THE DRAWINGS

In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:

Fig. 1 is a system block diagram of the present invention;

FIG2 is an illustrative example diagram of the semantic enhancement sampling strategy.

Detailed ways

The technical solution of the present invention is described in detail below with reference to the accompanying drawings.

The semantically enhanced graph contrast learning framework for road network representation provided by the present invention is used for road network representation learning. At present, the increasing complexity of urban traffic networks has led to massive and diverse road network data, and the processing and analysis of these data have brought huge storage and computing overhead. At present, road network data are usually processed by combining road attributes and visual information. The present invention proposes a new framework for the characterization and optimization of these data, including the implementation of the following four stages: first, a multimodal feature embedding module is constructed to obtain the attribute and visual feature embedding of the road segment using the attribute data of the road segment and the street view image data; second, through the graph augmentation module, two augmentation strategies are designed to perturb the topology and node features of the road network to generate two different augmented graphs; then, in the graph representation learning module, a graph neural network model is used to extract the representation of the road segment from the two graph variants; finally, in the semantically enhanced contrast optimization module, sample pairs are sampled in combination with the road network characteristics and the mobile semantics in the historical trajectory, and a dual-view contrast loss function is designed to guide the optimization process. This framework improves the efficiency of road network data processing and the understanding of the complex dynamics of urban traffic networks. Figure 1 is a system block diagram of the present invention.

In this embodiment, seven necessary concepts are involved, which are as follows:

The first concept: road network, a road network is a directed graph G = (V, A). Each vertex v∈V represents a section of road, which contains attributes such as length and coordinates on the map. A is an adjacency matrix, in which each entry Ai _,j is a binary value indicating whether the road segment _vj is connected to the road segment _vi .

Concept 2: Street View Imagery A panoramic image is a picture taken by a vehicle at a specific location on the road network. Specifically, we collect multiple images of each road segment and stitch them into a 360-degree panoramic view. A panoramic image represents a snapshot that illustrates the visual details of the surrounding environment.

Concept 3: Trajectory is a sequence of GPS points recorded when the vehicle moves on the road network, denoted as <p ₁ ,p ₂ ,p ₃ ,p ₄ ,...,p _|T| >. Where p _i is the coordinate of the i-th point and |T| represents the total number of points. We further map each trajectory onto the road network using our proposed L2MM deep model. Therefore, a trajectory is represented by a series of connected road segments <v ₁ ,v ₂ ,v ₃ ,v ₄ ,...,v _m > in the road network G.

The fourth concept, z _i, refers to the embedding or representation of a specific node (such as a road segment). In a graph, each node has a unique embedding to represent its characteristics. Represents the embedding of the positive sample. In contrastive learning, the positive sample is a node that is similar or related to the anchor node (z _i ). /> Represents the embedding of negative samples. In contrast to positive samples, negative samples are nodes that are dissimilar or irrelevant to the anchor node.

The fifth concept, anchor node, refers to a reference point or target node for comparison. During the learning process, the representation (or feature vector) of the anchor node is usually compared with the representation of other nodes (positive samples or negative samples) to learn discriminative node embeddings.

The sixth concept, h-hop, is used to describe the neighborhood of a node in a graph. In a graph, the h-hop neighborhood of a node refers to the set of all nodes that can be reached from this node by no more than h hops.

The seventh concept, InfoNCE objective, is a loss function commonly used in contrastive learning. The basic idea is to maximize the similarity between positive pairs and minimize the similarity between negative pairs. In contrastive learning, positive pairs are usually similar or related entities, while negative pairs are unrelated entities.

FIG2 is an illustrative example diagram of the semantic enhancement sampling strategy. The system framework of the present system mainly includes four modules: a multimodal feature embedding module, a graph augmentation module, a graph representation learning module, and a semantic enhancement-based contrast optimization module. FIG1 shows a system block diagram of the present invention, wherein: a multimodal feature embedding module: The starting point of this module is to use the attribute data of the road segment and the street view image data to obtain the attribute embedding and visual feature embedding of the road segment respectively, and then combine them together as the initial embedding of the road segment, including the following steps:

Step 1: Input road network G = (V, A) and street view image data And historical GPS track data/> The attribute embedding ^Xa and visual feature embedding ^Xv of the road segment are obtained through the multimodal feature embedding module;

Step 2: Concatenate the attribute embedding and visual feature embedding of the road segment to construct the initial embedding X = ^Xa ⊕ ^Xv of the road segment, and use it as the node (road segment) feature matrix of the road network.

Graph augmentation module: This module uses two different augmentation strategies to perturb the topology and node features of the road network (graph) respectively, thereby generating two different augmented graphs (graph variants), including the following steps:

Step 1: Calculate edge deletion probability based on road history trajectory data where e _i,j represents the edge connection between segments _vi and v _j , /> represents the estimated transition probability from segment vi _to segment _vj ;

Step 2: Based on the edge deletion probability Pr and the road network G as input, perturb the road network twice (each perturbation involves two aspects: topology and node features), and obtain the first augmented graph Its adjacency matrix is/> The feature matrix is/> And the second augmented image/> Its adjacency matrix is/> The feature matrix is/>

Graph representation learning module: The starting point of this module is to use the graph neural network model to extract the node level, that is, the representation of the road segment, based on the two graph variants generated by the graph augmentation module. It includes the following steps:

Step 1: Use the graph encoder to encode the adjacency matrix of the first augmented graph obtained in the graph augmentation module and feature matrix/> Perform encoding operation to obtain the embedding matrix/>

Step 2: Use the graph encoder to encode the adjacency matrix of the second augmented graph obtained in the graph augmentation module and feature matrix/> Perform encoding operation to obtain the embedding matrix/>

Step 3: Through the nonlinear projection head, the two embedding matrices obtained by the graph encoder are mapped to the low-dimensional latent space to obtain the embeddings Z ₁ and Z ₂ of the two groups of road segments.

Contrastive optimization module based on semantic enhancement: The starting point of this module is to sample sample pairs using the characteristics of the road network and the mobility semantics in the historical trajectory, and to design a dual-view contrast loss function to guide the contrast optimization process and use this loss function to train the network to optimize the road network embedding representation, including the following steps:

Step 1: Construct a co-occurrence matrix M based on the road history trajectory T. Then, based on the co-occurrence matrix M and the number of neighbor nodes h, the adjacency matrices of the two augmented graphs are respectively and/> Sampling of positive and negative samples. Specifically, given a target node (i.e., anchor node) in a certain view, its neighbor nodes within h hops are considered as intra-view positive samples, and the mirror nodes of another view are considered as inter-view positive samples. For nodes beyond h hops, further exclude nodes that co-occur with the anchor node in the historical trajectory, and select the remaining nodes as true negative samples;

Step 2: Based on the sampling results of step 1 as input, use the dual-view contrast loss function designed based on the InfoNCE target Calculate the loss value and use it as a guide to train the network.

Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims (6)

1. A semantic enhancement map contrast learning method for road network representation is characterized in that: the method constructs a semantic enhancement graph contrast learning framework for road network representation, and specifically comprises the following stages:

The first stage: constructing a multi-mode feature embedding module, respectively obtaining attribute embedding and visual feature embedding of the road section by utilizing the attribute data and street view image data of the road section, and splicing the attribute embedding and the visual feature embedding of the road section to serve as initial embedding of the road section;

and a second stage: constructing a graph augmentation module, designing two different augmentation strategies on the basis of modeling a road network as a graph, and respectively applying disturbance to topology and node characteristics of the road network so as to generate two different augmentation graphs;

And a third stage: constructing a graph representation learning module, and extracting node levels, namely the representation of road sections, by adopting a graph neural network model on the basis of two graph varieties generated by the graph augmentation module;

fourth stage: a contrast optimization module based on semantic enhancement is constructed, sampling of sample pairs is carried out by utilizing road network characteristics and mobile semantics in a historical track, and a contrast loss function of double views is designed to guide a contrast optimization process.

2. A semantic enhanced graph contrast learning method for road network representation according to claim 1, characterized by: in the first stage, the method mainly comprises attribute feature embedding and visual feature embedding;

The attribute feature embedding section includes: for any road section v _i, the road section identifier ID, the road section length LEN and the road section midpoint coordinate (LON, LAT) attributes are integrated for analysis, and the signs corresponding to the attributes are v _i.id、v_i.len、v_i.lon、v_i and LAT respectively; for real type attributes v _i. Lon and v _i. Lat, converting them into a discretized form through a binning operation, then mapping each attribute value into a corresponding feature vector using a separate linear layer, and then stitching together the four feature vectors to form the attribute embeddings of the road segments;

the visual characteristic embedding section includes: the pre-trained Swin transducer is used as a visual characteristic encoder, panoramic image data collected at the midpoint of a road section is used as input, and visual embedding of the panoramic image data is output Finally, two different feature types are connected together to obtain the initial embedding of the road segment.

3. A semantic enhanced graph contrast learning method for road network representation according to claim 2, characterized by: in the second stage, a graph augmentation module is constructed to augment the road network by using two strategies of mobility-based edge removal and modality-based feature coverage; the mobility-based edge removal strategy aims at the topological structure of the road network, and the edge removal probability is determined according to the transition probability among road sections in the historical track data; if two road segments occur together often in a historical track, the connection between them is considered strong and important, so the probability of removing such a connection will be low; the feature masking strategy based on the mode is a disturbance strategy aiming at node features, and one of the attributes or the image features is randomly selected to mask so that the contrast learning model is adapted to the loss of road section data, and therefore the robustness of the model is enhanced; during each training process, two image views are generated simultaneously for comparison by using the two different augmentation strategies.

4. A semantic enhanced graph contrast learning method for road network representation according to claim 3, characterized by: in the third stage, a graph representation learning module is constructed, and the two graph views generated by the graph augmentation module are taken as inputs to respectively learn the representation of the two graph views; the method specifically comprises the following steps: advanced features are learned and extracted from the graph view using a graph encoder to generate a complex representation of each segment, and finally, the segment representations generated by the graph encoder are mapped to a low-dimensional potential space by a nonlinear projection head.

5. The semantic enhanced graph contrast learning method for road network representation according to claim 4, wherein: in the fourth stage, a contrast optimization module based on semantic enhancement is constructed, wherein the module comprises a semantic enhancement sampling strategy, and positive and negative samples can be constructed by utilizing road network characteristics and mobile semantics in a historical track; the method specifically comprises the following steps: giving a target node of a certain view, regarding a neighbor node in the h-hop as an intra-view positive sample, and regarding a mirror node of another view as an inter-view positive sample; for nodes beyond the h-hop range, further screening out nodes co-occurring with anchor nodes in the history track, and selecting the rest nodes as real negative samples; a contrast loss function based on InfoNCE target design was then used:

to calculate the contrast loss of all nodes on the two graph views and obtain the final dual-view contrast loss function

Wherein v' _i represents the mirror node of v _i in another view; finally, training of the network is performed using the loss function to optimize the road network embedded representation.

6. A semantic enhanced graph contrast learning system for road network representation, characterized by: the system employs the method of any one of claims 1 to 5.