Methods, devices and electronic equipment for generating fake tags

Abstract

The present application is applicable to the technical field of intelligent transportation systems, and provides a pseudo-label generation method, a pseudo-label generation device, and an electronic device. The method comprises: obtaining graph data; if there is at least one label sparse subgraph and at least one label dense subgraph in the graph data, determining a matching target label dense subgraph and a target label sparse subgraph from the at least one label sparse subgraph and the at least one label dense subgraph; based on the real labels of the nodes in the target label dense subgraph, generating pseudo-labels of the nodes in the target label sparse subgraph. The present application can effectively increase the labels of the training data set.

Description

Technical Field

[0001] The present application belongs to the technical field of intelligent transportation systems, and in particular, relates to a pseudo-label generation method, a pseudo-label generation device, and an electronic device. Background Art

[0002] Traffic status recognition is one of the core functions of intelligent transportation systems. New detection equipment (such as radar video integrated machine) can collect information such as traffic flow, queues, speed and vehicle trajectory of each lane at the intersection, and then identify the traffic status of the lane (smooth, slow, congested, etc.). However, in the actual road network, there are still a large number of intersections equipped with traditional detectors. Traditional detectors can only collect basic information such as traffic flow and cannot accurately identify the road traffic status.

[0003] Some studies have attempted to use data-driven methods to identify traffic conditions: using basic information such as map data, traffic flow, and traffic light status as input features, and using the corresponding traffic status as labels to train a traffic status classification model. In the above methods, there are two main ways to obtain labels: (1) identifying traffic conditions through new detection equipment; (2) manually calibrating traffic conditions through real videos. In practical applications, due to the uneven distribution of new detection equipment and the high cost of manual calibration, it is still difficult to obtain sufficient labels, resulting in the inability to effectively train traffic status classification models. Therefore, before training a traffic status classification model, how to effectively increase the labels of the training data set is a technical problem that needs to be solved urgently. Summary of the invention

[0004] The embodiments of the present application provide a pseudo-label generation method, a pseudo-label generation device, and an electronic device, which can effectively increase the labels of a training data set.

[0005] In a first aspect, an embodiment of the present application provides a pseudo label generation method, comprising:

[0006] Acquire graph data; the nodes in the graph data are traffic flow directions of the traffic network, the edges in the graph data represent the connection relationship between the traffic flow directions, and the labels of the nodes in the graph data represent the traffic status;

[0007] If there is at least one labeled sparse subgraph and at least one labeled dense subgraph in the graph data, determine a matching target labeled dense subgraph and a target labeled sparse subgraph from the at least one labeled sparse subgraph and the at least one labeled dense subgraph; the labeled sparse subgraph is a subgraph in the graph data whose label density is less than a first density threshold, and the labeled dense subgraph is a subgraph in the graph data whose label density is greater than a second density threshold, and the first density threshold is less than the second density threshold;

[0008] Based on the real labels of the nodes in the target label dense subgraph, pseudo labels of the nodes in the target label sparse subgraph are generated.

[0009] In an embodiment of the present application, by obtaining graph data with traffic flow directions of a traffic network as nodes, connection relationships between traffic flows as edges, and node labels representing traffic states, it is possible to determine a matching target label dense subgraph and a target label sparse subgraph from at least one label sparse subgraph and at least one label dense subgraph when there is at least one label sparse subgraph and at least one label dense subgraph in the graph data. Since the target label dense subgraph and the target label sparse subgraph match, matching nodes in the target label dense subgraph and the target label sparse subgraph usually have the same label information. Therefore, based on the true labels of the nodes in the target label dense subgraph, pseudo labels of the nodes in the target label sparse subgraph can be generated, that is, the true labels of the target label dense subgraph are migrated to the target label sparse subgraph, which can effectively increase the labels of the training data set and make the label distribution more even, so as to better train traffic state classification models such as graph neural networks, thereby improving the overall state recognition accuracy of the traffic network.

[0010] In a second aspect, an embodiment of the present application provides a pseudo-label generating device, including:

[0011] A graph data acquisition module, used to acquire graph data; the nodes in the graph data are traffic flow directions of a traffic network, the edges in the graph data represent the connection relationship between the traffic flow directions, and the labels of the nodes in the graph data represent the traffic status;

[0012] A target determination module, for determining a matching target label dense subgraph and a target label sparse subgraph from the at least one label sparse subgraph and the at least one label dense subgraph if there are at least one label sparse subgraph and at least one label dense subgraph in the graph data; the label sparse subgraph is a subgraph in the graph data whose label density is less than a first density threshold, the label dense subgraph is a subgraph in the graph data whose label density is greater than a second density threshold, and the first density threshold is less than the second density threshold;

[0013] The pseudo-label generation module is used to generate pseudo-labels of nodes in the target label sparse subgraph based on the real labels of the nodes in the target label dense subgraph.

[0014] In a third aspect, an embodiment of the present application provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the electronic device implements the pseudo-label generation method as described in the first aspect above.

[0015] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a computer, the pseudo-label generating method as described in the first aspect above is implemented.

[0016] In a fifth aspect, an embodiment of the present application provides a computer program product, including a computer program, wherein when the computer program is executed, the pseudo-label generating method as described in the first aspect above is executed.

[0017] It can be understood that the beneficial effects of the second to fifth aspects mentioned above can be found in the relevant description of the first aspect mentioned above, and will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0019] Figure 1 It is a flowchart of a pseudo-label generation method provided in an embodiment of the present application;

[0020] Figure 2 is an example diagram of a transportation network;

[0021] Figure 3 yes Figure 2 An example diagram of graph data of a transportation network shown;

[0022] Figure 4 It is a flowchart of a method for determining a target label dense subgraph and a target label sparse subgraph provided in an embodiment of the present application;

[0023] Figure 5 is a flow chart of a method for determining a target similarity error provided in an embodiment of the present application;

[0024] Figure 6 is an example diagram of the graph matching process provided in an embodiment of the present application;

[0025] Figure 7-1 is an example graph of graph data before pseudo-labeling;

[0026] Figure 7-2 Is for Figure 7-1 An example of graph data after pseudo-labeling of nodes without labels in the graph;

[0027] Figure 8 is a schematic diagram of the structure of a pseudo-label generating device provided in an embodiment of the present application;

[0028] Fig. 9 It is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0029] In the following description, specific details such as specific system structures, technologies, etc. are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present application. However, it should be clear to those skilled in the art that the present application may also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the present application.

[0030] It should be understood that when used in the present specification and the appended claims, the term "comprising" indicates the presence of described features, wholes, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or combinations thereof.

[0031] It should also be understood that the term “and / or” used in the specification and appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0032] As used in the specification and appended claims of this application, the term "if" can be interpreted as "when" or "uponce" or "in response to determining" or "in response to detecting", depending on the context. Similarly, the phrase "if it is determined" or "if [described condition or event] is detected" can be interpreted as meaning "uponce it is determined" or "in response to determining" or "uponce [described condition or event] is detected" or "in response to detecting [described condition or event]", depending on the context.

[0033] In addition, in the description of the present application specification and the appended claims, the terms "first", "second", "third", etc. are only used to distinguish the descriptions and cannot be understood as indicating or implying relative importance.

[0034] References to "one embodiment" or "some embodiments" etc. described in the specification of this application mean that one or more embodiments of the present application include specific features, structures or characteristics described in conjunction with the embodiment. Therefore, the statements "in one embodiment", "in some embodiments", "in some other embodiments", "in some other embodiments", etc. that appear in different places in this specification do not necessarily refer to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized in other ways. The terms "including", "comprising", "having" and their variations all mean "including but not limited to", unless otherwise specifically emphasized in other ways.

[0035] The pseudo-label generation method provided in the embodiment of the present application can be applied to electronic devices such as mobile phones, tablet computers, laptop computers, servers, ultra-mobile personal computers (UMPCs), netbooks, personal digital assistants (PDAs), etc. The embodiment of the present application does not impose any restrictions on the specific types of electronic devices.

[0036] See also Figure 1 , Figure 1 A flow chart of a pseudo-label generation method provided in an embodiment of the present application is shown. As an example but not a limitation, the method is applied to an electronic device, and the method includes the following steps:

[0037] Step 101, obtaining graph data.

[0038] Among them, the nodes in the graph data are the traffic flows of the traffic network, the edges in the graph data represent the connection relationship between the traffic flows, and the labels of the nodes in the graph data represent the traffic status.

[0039] The above graph data can be constructed by the electronic device based on the traffic network, or can be manually constructed and then input into the electronic device. This application does not limit the source of the above graph data. As an example but not a limitation, the electronic device can abstract the traffic network into graph data with traffic flow directions as nodes and the connection relationship between traffic flow directions as edges.

[0040] After acquiring the graph data, the electronic device can migrate the real labels of the densely labeled subgraph to the sparsely labeled subgraph, thereby marking pseudo-labels for the unlabeled nodes in the sparsely labeled subgraph, effectively increasing the labels of the training data set, and converting the traffic state recognition problem into a node classification problem, so as to make full use of the spatial correlation characteristics contained in the traffic network, and improve the recognition accuracy of the overall traffic state of the traffic network. It should be understood that the label of the node may include a real label and a pseudo-label. The real label of a node may refer to the label that the node itself has (such as a manually annotated label). The pseudo-label of a node may refer to a label generated based on the real label.

[0041] Among them, the above traffic flow directions include but are not limited to left turn, straight go, right turn, etc.

[0042] The attribute information of nodes in the graph data includes but is not limited to direction, turn (i.e. traffic flow direction), number of lanes, traffic flow, green light duration and signal light cycle duration, etc., which represent the operation characteristics of traffic flow. Directions include but are not limited to the four directions of southeast, northwest and northeast. Directions, turns and number of lanes represent the characteristics of network topology. Traffic flow represents the dynamic characteristics of traffic flow. Green light duration and signal light cycle duration represent the influence of traffic light constraints on traffic flow. Node categories (i.e. node labels) include but are not limited to traffic states such as congestion, slow movement and smooth flow.

[0043] like Figure 2 The following is an example diagram of a transportation network. Figure 3 Said is Figure 2 An example diagram of the graph data of the transportation network shown in FIG. Figure 3 The traffic flows of different directions (i.e. left turn, straight ahead, right turn) converge and disperse through the confluence point. At the same time, six-dimensional parameters such as direction, turn, number of lanes, traffic flow, green light duration, and cycle duration (i.e. signal light cycle duration) are selected to characterize the operation characteristics of traffic flow.

[0044] Step 102: If there is at least one labeled sparse subgraph and at least one labeled dense subgraph in the graph data, determine a matching target labeled dense subgraph and target labeled sparse subgraph from the at least one labeled sparse subgraph and the at least one labeled dense subgraph.

[0045] Among them, the label sparse subgraph is a subgraph in the graph data whose label density is less than the first density threshold, and the label dense subgraph is a subgraph in the graph data whose label density is greater than the second density threshold, and the first density threshold is less than the second density threshold.

[0046] It should be understood that the number of nodes in the target label dense subgraph may be greater than the number of nodes in the target label sparse subgraph, may be less than the number of nodes in the target label sparse subgraph, or may be equal to the number of nodes in the target label sparse subgraph.

[0047] It should be noted that the nodes in the matching target label dense subgraph are matched one-to-one with the nodes in the target label sparse subgraph. The one-to-one matching can ensure that the nodes in the subgraph with a small number of nodes have a corresponding matching node in the subgraph with a large number of nodes.

[0048] In one embodiment, the electronic device may first determine at least two subgraphs included in the graph data, and then determine whether there are label sparse subgraphs and label dense subgraphs in the at least two subgraphs based on the first density threshold and the second density threshold, and if there are at least one label sparse subgraph and at least one label dense subgraph, perform step 102. Optionally, the first density threshold and the second density threshold may be set according to scenario requirements.

[0049] In a possible implementation, a k-order neighbor subgraph of each node in the graph data may be determined as a subgraph of the graph data, where k is an integer greater than 0. Optionally, k may be set according to scenario requirements, for example, k is 2.

[0050] On the basis of determining the k-order neighbor subgraph of each node in the graph data as the subgraph of the graph data, the method of determining the label sparse subgraph and the label dense subgraph includes:

[0051] Determine the total number of nodes in the k-order neighbor subgraph of each node in the graph data and the number of nodes marked with true labels;

[0052] For any node's k-order neighbor subgraph, determine the label density of the k-order neighbor subgraph based on the total number of nodes in the k-order neighbor subgraph and the number of nodes marked with true labels;

[0053] If the label density of the k-order neighbor subgraph is less than the first density threshold, the k-order neighbor subgraph is determined to be a label sparse subgraph;

[0054] If the label density of the k-order neighbor subgraph is greater than the second density threshold, the k-order neighbor subgraph is determined to be a label-dense subgraph.

[0055] For any node's k-order neighbor subgraph, the number of nodes marked with true labels in the k-order neighbor subgraph is divided by the total number of nodes in the k-order neighbor subgraph, and the resulting value is determined as the label density of the k-order neighbor subgraph.

[0056] As an example and not a limitation, assume that the traffic network is defined as a graph data .in represents a set of N nodes, is an edge set, through The adjacency matrix can be constructed , represents the number of nodes in the adjacency matrix, is the attribute matrix of the node, each node There is a unique dimensional feature vector .also, represents the true label set, is the number of categories of nodes, Is a node The real label.

[0057] In the labeled sparse subgraph, a small set of nodes annotated with true labels is represented as , and its corresponding label set is At the same time, there are still a large number of unlabeled nodes in the label sparse subgraph, and their set is represented as ,at this time .

[0058] For graph data , the label density of each node (i.e., the label density of the k-order neighbor subgraph of the node) is defined as shown in formula (1).

[0059]

[0060] in, Representation Node The number of nodes marked with true labels in the k-order neighbor subgraph, Representation Node The total number of nodes in the k-order neighbor subgraph, It indicated The ratio of the number of nodes marked with true labels to the total number of nodes in the k-order neighbor subgraph. The closer it is to 1, otherwise it is close to 0. Based on formula (1), the label density of all nodes can be calculated and the first density threshold can be set respectively. and the second density threshold , which is used to filter out label sparse subgraphs and label dense subgraphs from the k-order neighbor subgraphs of all nodes. The set of label sparse subgraphs can be expressed as , Is a node The k-order neighbor subgraph satisfies Similarly, the set of labeled dense subgraphs can be represented as , Is a node The k-order neighbor subgraph satisfies .

[0061] See also Figure 4 In a possible implementation, the electronic device may determine a target label dense subgraph and a target label sparse subgraph according to steps 401 to 402.

[0062] Step 401, determining the target similarity errors of the p*q sub-graph combinations.

[0063] Among them, the target similarity error represents the difference between the sparsely labeled subgraph and the densely labeled subgraph in the corresponding subgraph combination under the target node matching mode. The target node matching mode represents the node matching relationship that minimizes the similarity error of the corresponding subgraph combination. The node matching relationship is the matching relationship between the nodes in the sparsely labeled subgraph and the nodes in the densely labeled subgraph in the corresponding subgraph combination. p is the number of sparsely labeled subgraphs, and q is the number of densely labeled subgraphs.

[0064] In label sparse collection The Chinese Communist Party sparsely labeled subgraphs, densely labeled sets The Chinese Communist Party In the case of a labeled dense subgraph, a labeled sparse subgraph and a labeled dense subgraph that need to be matched are called a subgraph combination, then labeled sparse subgraphs and When the labeled dense subgraphs are matched pairwise, there are p*q subgraph combinations.

[0065] For any subgraph combination, there is at least one node matching pattern between the labeled sparse subgraph and the labeled dense subgraph in the subgraph combination. Since the matching relationship between the nodes in the labeled sparse subgraph and the nodes in the labeled dense subgraph under different node matching patterns is different, the matching relationship affects the similarity error. Therefore, the similarity errors corresponding to different node matching patterns may be different. The smaller the similarity error, the more matched the labeled sparse subgraph and the labeled dense subgraph in the subgraph combination under the corresponding node matching pattern. Therefore, the target similarity error can be selected from the similarity errors corresponding to at least one node matching pattern of the subgraph combination, and the matching target labeled dense subgraph and target labeled sparse subgraph are determined based on the target similarity error.

[0066] In a possible implementation, the electronic device may determine the target similarity error of the sub-graph combination by graph matching. Specifically, for any label sparse sub-graph of the sub-graph combination, and labeled dense subgraph , a secondary optimization model is constructed using graph structure similarity and node attribute similarity, and the target similarity error of the subgraph combination is obtained by solving the secondary optimization model. Optionally, the graph structure error can be used to represent the graph structure similarity, and the smaller the graph structure error, the greater the graph structure similarity, and the larger the graph structure error, the smaller the graph structure similarity.

[0067] Graph matching refers to a pattern recognition method based on graph theory, which aims to find the optimal node matching relationship between two subgraphs. This node matching relationship can be completely isomorphic or an approximate match that meets specific constraints.

[0068] See also Figure 5 The scheme of determining the target similarity error of the sub-graph combination by graph matching includes steps 501 to 503.

[0069] Step 501 : for any subgraph combination among the p*q subgraph combinations, based on the permutation matrix corresponding to the node matching pattern of the subgraph combination, determine the graph structure error function and the node attribute similarity function of the subgraph combination in the corresponding node matching pattern.

[0070] Among them, the node matching mode can be represented by a permutation matrix. Since each node matching mode of the subgraph combination corresponds to a similarity error, determining the minimum similarity error is to determine the target node matching mode, and the permutation matrix is ​​an unknown quantity. The graph structure error function characterizes the degree of graph structure difference between the sparsely labeled subgraph and the densely labeled subgraph in the subgraph combination under the node matching mode.

[0071] Since the permutation matrix is ​​an unknown quantity, the graph structure error determined based on the permutation matrix can be called a graph structure error function, and the node attribute similarity determined based on the permutation matrix can be called a node attribute similarity function.

[0072] In a possible implementation, the graph structure error function is determined by:

[0073] Determine a first adjacency matrix and a second adjacency matrix; the first adjacency matrix is ​​an adjacency matrix of a sparsely labeled subgraph in the subgraph combination, and the second adjacency matrix is ​​an adjacency matrix of a densely labeled subgraph in the subgraph combination;

[0074] A graph structure error function is determined based on the permutation matrix, the first adjacency matrix, and the second adjacency matrix.

[0075] If two subgraphs have similar structural features, then their corresponding nodes may have the same label. On this basis, the first adjacency matrix can be used Represents a labeled sparse subgraph The structural features of the second adjacency matrix Represents a densely labeled subgraph The structural features are calculated in any node matching mode and The graph structure error function between is shown in formula (2).

[0076]

[0077] in, represents the F-norm, is the permutation matrix corresponding to the node matching pattern, and Represents the label sparse subgraphs and labeled dense subgraph The number of nodes. Represents a labeled sparse subgraph The i-th node and label dense subgraph The jth node of is matched (that is, there is a matching relationship), otherwise .

[0078] In a possible implementation, the node attribute similarity function is determined by:

[0079] Determine a first vector representation and a second vector representation; the first vector representation is a vector representation of a node in a sparsely labeled subgraph in the subgraph combination, and the second vector representation is a vector representation of a node in a densely labeled subgraph in the subgraph combination;

[0080] A node attribute similarity function is determined based on the permutation matrix, the first vector representation, and the second vector representation.

[0081] If two subgraphs have similar node attribute characteristics, then their corresponding nodes may have the same label. On this basis, we can calculate the node matching mode. and The node attribute similarity function between them is shown in formula (3).

[0082]

[0083] in, Represents a labeled sparse subgraph generated by a Graph Neural Network (GCN) model The vector representation of all nodes of Represents the labeled dense subgraph generated by the GCN model The vector representation of all nodes.

[0084] The GCN model is a deep learning model specifically used to process graph data. It can effectively capture the association relationship between nodes and learn and represent features based on this.

[0085] Step 502: Determine a similarity error function based on the graph structure error function and the node attribute similarity function.

[0086] This embodiment migrates label information through graph structure errors and node attribute similarities, which can effectively increase label density, make label distribution more uniform, improve the recognition accuracy of traffic status in label-sparse scenarios, and have better engineering applicability. The identified traffic status can assist traffic management departments in making decisions and taking effective management measures for related roads.

[0087] Step 503 , under the constraint that nodes in the sparsely labeled subgraphs match nodes in the densely labeled subgraphs in the subgraph combination one by one, solve the minimum value of the similarity error function, and determine the minimum value as the target similarity error of the subgraph combination.

[0088] The electronic device subtracts the node attribute similarity function from the graph structure error function to obtain the similarity error function, that is, to construct a secondary distribution model for solving the target node matching pattern, as shown in formula (4).

[0089]

[0090] in, represent dimensional column vector, All values ​​in are 1, and the above constraints ensure that the label sparse subgraph A dense subgraph of nodes and labels The nodes of are matched one to one. For example, when The above constraints can ensure Each node can be associated with a labeled dense subgraph The nodes of are matched one by one, and Not all nodes need to be included in the label sparse subgraph. The matching nodes are found in the query.

[0091] like Figure 6 The figure shows an example of the graph matching process provided by the embodiment of the present application. Assume that there is Figure 6 The labeled dense subgraph A1 and the labeled sparse subgraph A2 are shown. The four nodes of the labeled dense subgraph A1 are all labeled with real labels, and the three nodes of the labeled sparse subgraph A2 are all unlabeled. The node matching pattern of the labeled dense subgraph A1 and the labeled sparse subgraph A2 (that is, node 1 matches node 6, node 2 matches node 5, and node 3 matches node 7) can be represented by the permutation matrix Π, which can be expressed by Figure 6 The process shown in the figure calculates the graph structure error (i.e. Figure 6 The similarity of subgraph structures in Figure 6 subgraph attribute similarity description in ).

[0092] Step 402 : when the p*q target similarity errors are arranged in ascending order, the sub-graph combination corresponding to the target similarity errors arranged in the first w positions is determined as the target sub-graph combination.

[0093] Among them, the label dense subgraph in the target subgraph combination is the target label dense subgraph, and the label sparse subgraph in the target subgraph combination is the target label sparse subgraph.

[0094] When matching, each subgraph combination will determine the corresponding target node matching mode and target similarity error through formula (4). Then, p*q subgraph combinations correspond to p*q target similarity errors, that is, the matching results between all label sparse subgraphs and all label dense subgraphs are obtained. Then, the p*q target similarity errors are sorted in order from small to large, and the most matching w subgraph combinations are selected as the target subgraph combinations. Optionally, w can be set according to the scene requirements. For example, w can be 5% of p*q.

[0095] Step 103: Generate pseudo labels for nodes in the target label sparse subgraph based on the real labels of the nodes in the target label dense subgraph.

[0096] Through step 103, pseudo labels can be assigned to nodes in the target label sparse subgraph, thereby achieving full utilization of label information and increasing the label density of the training data set.

[0097] In a possible implementation, generating pseudo labels of nodes in a target label sparse subgraph based on real labels of nodes in a target label dense subgraph includes:

[0098] For any target node, if the node matching the target node in the target label dense subgraph is marked with a true label, a pseudo label of the target node is generated based on the true label of the node matching the target node in the target label dense subgraph; the target node is a node without a label in the target label sparse subgraph.

[0099] The node in the target label dense subgraph that matches the target node may refer to a node in the target label dense subgraph that matches or corresponds to the target node in a target node matching mode.

[0100] Since the target label dense subgraph and the target label sparse subgraph have similar graph structures and node attributes in the target node matching mode, the target node usually has the same label information as the node matching the target node in the target label dense subgraph, and the true label of the node matching the target node in the target label dense subgraph can be determined as the pseudo label of the target node.

[0101] In the p*q subgraph combination, one label sparse subgraph needs to be matched with q label dense subgraphs, that is, there is a target label sparse subgraph corresponding to at least two target label dense subgraphs. In this case, the target similarity errors of the target label sparse subgraph and each target subgraph combination corresponding to each target label dense subgraph can be compared, and the true label of the node in the target label dense subgraph corresponding to the minimum value of each target similarity error that matches the target node is determined as the pseudo label of the target node.

[0102] For any target node, if the node matching the target node in the target label dense subgraph is not labeled with the true label, the target node may not be labeled.

[0103] In a possible implementation, after generating pseudo labels for nodes in the target label sparse subgraph, the electronic device may further perform the following steps:

[0104] Mark the pseudo labels on the corresponding nodes in the target label sparse subgraph;

[0105] Based on the real label node set and the pseudo label node set in the graph data, the traffic state classification model is trained to obtain the trained traffic state classification model; the real label node set includes the attribute information and real labels of the first type of nodes, and the pseudo label node set includes the attribute information and pseudo labels of the second type of nodes. The first type of nodes are nodes marked with real labels in the graph data, and the second type of nodes are nodes marked with pseudo labels in the graph data;

[0106] The attribute information of the node to be identified is input into the trained traffic status classification model to obtain the traffic status of the node to be identified.

[0107] Among them, the above-mentioned marking of the pseudo-label on the corresponding node in the target label sparse subgraph may refer to: associating the pseudo-label with the corresponding node in the target label sparse subgraph. As an example and not limitation, if the pseudo-label of node A in the target label sparse subgraph represents congestion, then the pseudo-label representing congestion can be associated with node A.

[0108] The real label node set and pseudo label node set in the graph data constitute the training data set, which is used to train the traffic state classification model.

[0109] The node to be identified may be any node in the graph data obtained in step 101 (for example, a node in the graph data that is not marked with a real label or a pseudo label), or may be any node in another graph data, which comes from a different transportation network than the graph data obtained in step 101.

[0110] As an example but not a limitation, the graph data obtained in step 101 is constructed based on the traffic network of city A. For the sake of ease of distinction, the graph data can be referred to as graph data A. On this basis, through the embodiment of the present application, pseudo labels for nodes in the labeled sparse subgraph in graph data A can be generated, and a traffic state classification model can be trained based on the real label node set and the pseudo label node set in graph data A. After the traffic state classification model is trained, the attribute information of any node in graph data A can be input into the trained traffic state classification model to obtain the traffic state of any node in graph data A. The attribute information of any node in graph data B constructed based on the traffic network of city B can also be input into the trained traffic state classification model to obtain the traffic state of any node in graph data B.

[0111] It should be understood that the present application does not limit the specific model structure of the above-mentioned traffic status classification model. As an example but not a limitation, the above-mentioned traffic status classification model can be a GCN model. When training the GCN model, the pseudo-label node set and the real label node set can be input into the GCN model together, and the Adaptive Moment Estimation (Adam) algorithm is used for optimization. After multiple rounds of iterations, a trained GCN model can be obtained. By classifying the nodes through the trained GCN model, the traffic status of all nodes in the traffic network can be inferred.

[0112] As an example but not a limitation, in an application scenario, there are areas in the traffic network where new detection equipment is densely deployed, and there are also areas where new detection equipment is sparsely deployed. The areas where new detection equipment is sparsely deployed can only collect basic information such as traffic flow, and cannot accurately identify the road traffic status. At this time, through the embodiment of the present application, the traffic status labels (i.e., real labels) collected by the new detector can be migrated to the label sparse area (i.e., label sparse subgraph), pseudo labels are generated for the nodes in the label sparse area, and then the traffic status of the entire road network is obtained by training the traffic status classification model.

[0113] By way of example and not limitation, Figure 7-1 is an example graph of graph data before pseudo-labeling. Figure 7-2 Is for Figure 7-1 An example of graph data after pseudo-labeling of nodes without labels in the graph. Figure 7-1 and Figure 7-2 The blue dots in the figure represent unlabeled nodes (i.e., nodes that are neither labeled with true labels nor with pseudo labels), and the red dots represent nodes that are labeled with true labels. Figure 7-2 The yellow dots in the figure represent nodes marked with pseudo labels. Figure 7-1 and Figure 7-2By comparison, it can be seen that after the real labels are migrated to the label sparse area through the embodiment of the present application, the label density of the training data set used to train the traffic state classification model can be increased. On this basis, the trained traffic state classification model can have a higher recognition accuracy, thereby more accurately obtaining the overall traffic state of the road network.

[0114] In an embodiment of the present application, by obtaining graph data with traffic flow directions of a traffic network as nodes, connection relationships between traffic flows as edges, and node labels representing traffic status, a matching target label dense subgraph and a target label sparse subgraph can be determined from at least one label sparse subgraph and at least one label dense subgraph when there is at least one label sparse subgraph and at least one label dense subgraph in the graph data. Since the target label dense subgraph and the target label sparse subgraph match, matching nodes in the target label dense subgraph and the target label sparse subgraph usually have the same label information. Therefore, based on the true labels of the nodes in the target label dense subgraph, pseudo labels of the nodes in the target label sparse subgraph can be generated, that is, the true labels of the target label dense subgraph are migrated to the target label sparse subgraph, which can effectively increase the labels of the training data set and make the label distribution more even, thereby improving the recognition accuracy of the overall traffic status of the traffic network.

[0115] It should be understood that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0116] Corresponding to the pseudo-label generation method described in the above embodiment, Figure 8 A schematic diagram of the structure of a pseudo-label generating device provided in an embodiment of the present application is shown. For the sake of convenience of explanation, only the parts related to the embodiment of the present application are shown.

[0117] Reference Figure 8 , the device comprises:

[0118] The graph data acquisition module 801 is used to acquire graph data; the nodes in the graph data are traffic flows of the traffic network, the edges in the graph data represent the connection relationship between the traffic flows, and the labels of the nodes in the graph data represent the traffic status;

[0119] A target determination module 802 is used to determine a matching target label dense subgraph and a target label sparse subgraph from the at least one label sparse subgraph and the at least one label dense subgraph if there is at least one label sparse subgraph and at least one label dense subgraph in the graph data; the label sparse subgraph is a subgraph in the graph data whose label density is less than a first density threshold, and the label dense subgraph is a subgraph in the graph data whose label density is greater than a second density threshold, and the first density threshold is less than the second density threshold;

[0120] The pseudo-label generation module 803 is used to generate pseudo-labels of nodes in the target label sparse subgraph based on the real labels of the nodes in the target label dense subgraph.

[0121] Optionally, the target determination module 802 includes:

[0122] A first determination unit is used to determine the target similarity error of each of the p*q subgraph combinations; the target similarity error represents the degree of difference between the label sparse subgraph and the label dense subgraph in the corresponding subgraph combination under the target node matching mode, the target node matching mode represents the node matching relationship that minimizes the similarity error of the corresponding subgraph combination, the node matching relationship is the matching relationship between the nodes in the label sparse subgraph and the nodes in the label dense subgraph in the corresponding subgraph combination, p is the number of the label sparse subgraphs, p is an integer greater than zero, q is the number of the label dense subgraphs, q is an integer greater than zero;

[0123] The second determination unit is used to determine the subgraph combination corresponding to the target similarity errors arranged in the first w positions as the target subgraph combination when p*q target similarity errors are arranged in ascending order; the label dense subgraph in the target subgraph combination is the target label dense subgraph, and the label sparse subgraph in the target subgraph combination is the target label sparse subgraph, and w is an integer greater than zero.

[0124] Optionally, the first determining unit includes:

[0125] A function determination subunit is used to determine, for any of the p*q subgraph combinations, a graph structure error function and a node attribute similarity function of the subgraph combination under the corresponding node matching mode based on a permutation matrix corresponding to the node matching mode of the subgraph combination; the permutation matrix is ​​an unknown quantity, and the graph structure error function represents the degree of difference in graph structure between a sparsely labeled subgraph and a densely labeled subgraph in the subgraph combination under the node matching mode;

[0126] An error determination subunit, configured to determine a similarity error function based on the graph structure error function and the node attribute similarity function;

[0127] The function solving subunit is used to solve the minimum value of the similarity error function under the constraint that the nodes in the sparsely labeled subgraph and the nodes in the densely labeled subgraph in the subgraph combination match one by one, and determine the minimum value as the target similarity error of the subgraph combination.

[0128] Optionally, the above function determines that the subunit is specifically used for:

[0129] Determine a first adjacency matrix and a second adjacency matrix; the first adjacency matrix is ​​an adjacency matrix of a sparsely labeled subgraph in the subgraph combination, and the second adjacency matrix is ​​an adjacency matrix of a densely labeled subgraph in the subgraph combination;

[0130] The graph structure error function is determined based on the permutation matrix, the first adjacency matrix, and the second adjacency matrix.

[0131] Optionally, the above function determines that the subunit is specifically used for:

[0132] Determine a first vector representation and a second vector representation; the first vector representation is a vector representation of a node in a sparsely labeled subgraph in the subgraph combination, and the second vector representation is a vector representation of a node in a densely labeled subgraph in the subgraph combination;

[0133] The node attribute similarity function is determined based on the permutation matrix, the first vector representation, and the second vector representation.

[0134] Optionally, the pseudo-label generating module 803 is specifically used for:

[0135] For any target node, if the node matching the target node in the target label dense subgraph is marked with a true label, a pseudo label of the target node is generated based on the true label of the node matching the target node in the target label dense subgraph; the target node is a node without a label in the target label sparse subgraph.

[0136] Optionally, the above device further includes:

[0137] A quantity determination module, used to determine the total number of nodes in the k-order neighbor subgraph of each node in the graph data and the number of nodes marked with real labels; k is an integer greater than zero;

[0138] A density determination module, configured to determine, for the k-order neighbor subgraph of any node, a label density of the k-order neighbor subgraph based on the total number of nodes in the k-order neighbor subgraph and the number of nodes marked with real labels;

[0139] A sparse determination module, configured to determine that the k-order neighbor subgraph is the label sparse subgraph if the label density of the k-order neighbor subgraph is less than the first density threshold;

[0140] A dense determination module is used to determine that the k-order neighbor subgraph is the label dense subgraph if the label density of the k-order neighbor subgraph is greater than the second density threshold.

[0141] It should be noted that the information interaction, execution process, etc. between the above-mentioned devices / units are based on the same concept as the method embodiment of the present application. Their specific functions and technical effects can be found in the method embodiment part and will not be repeated here.

[0142] Fig. 9 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application. Fig. 9 As shown, the electronic device 9 of this embodiment includes: at least one processor 90 ( Fig. 9 Only one is shown in the figure), a memory 91, and a computer program 92 stored in the memory 91 and executable on the at least one processor 90, wherein the processor 90 implements the steps of any of the above-mentioned method embodiments when executing the computer program 92.

[0143] The electronic device may include, but is not limited to, a processor 90 and a memory 91. Those skilled in the art will appreciate that Fig. 9 It is only an example of the electronic device 9 and does not constitute a limitation on the electronic device 9. It may include more or fewer components than shown in the figure, or a combination of certain components, or different components. For example, it may also include input and output devices, network access devices, etc.

[0144] The processor 90 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc.

[0145] In some embodiments, the memory 91 may be an internal storage unit of the electronic device 9, such as a hard disk or memory of the electronic device 9. In other embodiments, the memory 91 may also be an external storage device of the electronic device 9, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc. equipped on the electronic device 9. Further, the memory 91 may also include both an internal storage unit of the electronic device 9 and an external storage device. The memory 91 is used to store an operating system, an application program, a boot loader (BootLoader), data, and other programs, such as the program code of the computer program, etc. The memory 91 may also be used to temporarily store data that has been output or is to be output.

[0146] The technicians in the relevant field can clearly understand that for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In practical applications, the above-mentioned function allocation can be completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated in a processing unit, or each unit can exist physically separately, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, which will not be repeated here.

[0147] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present application implements all or part of the processes in the above-mentioned embodiment method, which can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a computer-readable storage medium. When the computer program is executed by the processor, the steps of the above-mentioned various method embodiments can be implemented. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may at least include: any entity or device that can carry the computer program code to the device / electronic device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electric carrier signal, a telecommunication signal, and a software distribution medium. For example, a USB flash drive, a mobile hard disk, a magnetic disk or an optical disk. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electric carrier signals and telecommunication signals.

[0148] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0149] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0150] In the embodiments provided in the present application, it should be understood that the disclosed devices / electronic devices and methods can be implemented in other ways. For example, the device / electronic device embodiments described above are merely schematic. For example, the division of the modules or units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0151] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0152] The embodiments described above are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, a person skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application, and should all be included in the protection scope of the present application.

Claims

1. A pseudo-label generation method, characterized in that: include: Get graph data; The nodes in the graph data are traffic flows of the traffic network, the edges in the graph data represent the connection relationship between the traffic flows, and the labels of the nodes in the graph data represent the traffic status; If there is at least one labeled sparse subgraph and at least one labeled dense subgraph in the graph data, determine a matching target labeled dense subgraph and a target labeled sparse subgraph from the at least one labeled sparse subgraph and the at least one labeled dense subgraph; the labeled sparse subgraph is a subgraph in the graph data whose label density is less than a first density threshold, and the labeled dense subgraph is a subgraph in the graph data whose label density is greater than a second density threshold, and the first density threshold is less than the second density threshold; Generate pseudo labels for nodes in the target label sparse subgraph based on the real labels for the nodes in the target label dense subgraph; The step of determining a matching target label dense subgraph and a target label sparse subgraph from the at least one label sparse subgraph and the at least one label dense subgraph comprises: Determine the target similarity error of each of the p*q subgraph combinations; the target similarity error represents the degree of difference between the label sparse subgraph and the label dense subgraph in the corresponding subgraph combination under the target node matching mode, the target node matching mode represents the node matching relationship that minimizes the similarity error of the corresponding subgraph combination, the node matching relationship is the matching relationship between the nodes in the label sparse subgraph and the nodes in the label dense subgraph in the corresponding subgraph combination, p is the number of the label sparse subgraphs, p is an integer greater than zero, q is the number of the label dense subgraphs, q is an integer greater than zero; When p*q target similarity errors are arranged in ascending order, the subgraph combination corresponding to the target similarity errors arranged in the first w positions is determined as the target subgraph combination; the label dense subgraph in the target subgraph combination is the target label dense subgraph, the label sparse subgraph in the target subgraph combination is the target label sparse subgraph, and w is an integer greater than zero; The determining of the target similarity errors of the p*q subgraph combinations includes: For any of the p*q subgraph combinations, based on the permutation matrix corresponding to the node matching mode of the subgraph combination, determine the graph structure error function and the node attribute similarity function of the subgraph combination under the corresponding node matching mode; the permutation matrix is ​​an unknown quantity, and the graph structure error function represents the degree of difference in graph structure between the label sparse subgraph and the label dense subgraph in the subgraph combination under the node matching mode; Determining a similarity error function based on the graph structure error function and the node attribute similarity function; Under the constraint that nodes in the sparsely labeled subgraph and nodes in the densely labeled subgraph in the subgraph combination are matched one by one, a minimum value of the similarity error function is solved, and the minimum value is determined as a target similarity error of the subgraph combination.

2. The pseudo-label generation method according to claim 1, characterized in that: The method for determining the graph structure error function includes: Determine a first adjacency matrix and a second adjacency matrix; the first adjacency matrix is ​​an adjacency matrix of a sparsely labeled subgraph in the subgraph combination, and the second adjacency matrix is ​​an adjacency matrix of a densely labeled subgraph in the subgraph combination; The graph structure error function is determined based on the permutation matrix, the first adjacency matrix, and the second adjacency matrix.

3. The pseudo-label generation method according to claim 1, characterized in that: The method for determining the node attribute similarity function includes: Determine a first vector representation and a second vector representation; the first vector representation is a vector representation of a node in a sparsely labeled subgraph in the subgraph combination, and the second vector representation is a vector representation of a node in a densely labeled subgraph in the subgraph combination; The node attribute similarity function is determined based on the permutation matrix, the first vector representation, and the second vector representation.

4. The pseudo-label generation method according to any one of claims 1 to 3, characterized in that: The generating of pseudo labels of nodes in the target label sparse subgraph based on the real labels of the nodes in the target label dense subgraph includes: For any target node, if the node matching the target node in the target label dense subgraph is marked with a true label, a pseudo label of the target node is generated based on the true label of the node matching the target node in the target label dense subgraph; the target node is a node without a label in the target label sparse subgraph.

5. The pseudo-label generation method according to any one of claims 1 to 3, characterized in that: The method further comprises: Determine the total number of nodes in the k-order neighbor subgraph of each node in the graph data and the number of nodes marked with real labels; k is an integer greater than zero; For the k-order neighbor subgraph of any node, determining the label density of the k-order neighbor subgraph based on the total number of nodes in the k-order neighbor subgraph and the number of nodes marked with real labels; If the label density of the k-order neighbor subgraph is less than the first density threshold, determining that the k-order neighbor subgraph is the label sparse subgraph; If the label density of the k-order neighbor subgraph is greater than the second density threshold, the k-order neighbor subgraph is determined to be the label-dense subgraph.

6. The pseudo-label generation method according to any one of claims 1 to 3, characterized in that: After generating pseudo labels of nodes in the target label sparse subgraph, the method further includes: Marking the pseudo labels on corresponding nodes in the target label sparse subgraph; Based on the real label node set and the pseudo label node set in the graph data, a traffic state classification model is trained to obtain a trained traffic state classification model; the real label node set includes attribute information and real labels of the first type of nodes, and the pseudo label node set includes attribute information and pseudo labels of the second type of nodes, the first type of nodes are nodes marked with real labels in the graph data, and the second type of nodes are nodes marked with pseudo labels in the graph data; The attribute information of the node to be identified is input into the trained traffic status classification model to obtain the traffic status of the node to be identified.

7. A pseudo-label generating device, characterized in that: include: A graph data acquisition module, used to acquire graph data; the nodes in the graph data are traffic flow directions of a traffic network, the edges in the graph data represent the connection relationship between the traffic flow directions, and the labels of the nodes in the graph data represent the traffic status; A target determination module, for determining a matching target label dense subgraph and a target label sparse subgraph from the at least one label sparse subgraph and the at least one label dense subgraph if there are at least one label sparse subgraph and at least one label dense subgraph in the graph data; the label sparse subgraph is a subgraph in the graph data whose label density is less than a first density threshold, the label dense subgraph is a subgraph in the graph data whose label density is greater than a second density threshold, and the first density threshold is less than the second density threshold; A pseudo-label generation module, used to generate pseudo-labels of nodes in the target label sparse subgraph based on the real labels of the nodes in the target label dense subgraph; The target determination module comprises: A first determination unit is used to determine the target similarity error of each of the p*q subgraph combinations; the target similarity error represents the degree of difference between the label sparse subgraph and the label dense subgraph in the corresponding subgraph combination under the target node matching mode, the target node matching mode represents the node matching relationship that minimizes the similarity error of the corresponding subgraph combination, the node matching relationship is the matching relationship between the nodes in the label sparse subgraph and the nodes in the label dense subgraph in the corresponding subgraph combination, p is the number of the label sparse subgraphs, p is an integer greater than zero, q is the number of the label dense subgraphs, q is an integer greater than zero; The second determining unit is used to determine the subgraph combination corresponding to the target similarity errors arranged in the first w positions as the target subgraph combination when p*q target similarity errors are arranged in order from small to large; the label dense subgraph in the target subgraph combination is the target label dense subgraph, and the label sparse subgraph in the target subgraph combination is the target label sparse subgraph, and w is an integer greater than zero; The first determining unit includes: A function determination subunit is used to determine, for any of the p*q subgraph combinations, a graph structure error function and a node attribute similarity function of the subgraph combination under the corresponding node matching mode based on a permutation matrix corresponding to the node matching mode of the subgraph combination; the permutation matrix is ​​an unknown quantity, and the graph structure error function represents the degree of difference in graph structure between a sparsely labeled subgraph and a densely labeled subgraph in the subgraph combination under the node matching mode; An error determination subunit, configured to determine a similarity error function based on the graph structure error function and the node attribute similarity function; The function solving subunit is used to solve the minimum value of the similarity error function under the constraint that the nodes in the sparsely labeled subgraph and the nodes in the densely labeled subgraph in the subgraph combination match one by one, and determine the minimum value as the target similarity error of the subgraph combination.

8. 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 computer program, the electronic device implements the pseudo-label generating method according to any one of claims 1 to 6.

Images