Uncertainty perception cross-graph learning method for truth value inference of conflict data in vehicle crowdsourcing

Abstract

The invention belongs to the field of intelligent traffic systems, and relates to an uncertainty perception cross-graph learning method for conflict data truth value inference in vehicle crowdsourcing, which comprises the following steps: obtaining independent views formed by different observation view angles corresponding to perception tasks; calculating attention coefficients among nodes in the independent views, and superposing attention layers of the views to obtain potential features; based on the potential features, calculating the similarity of every two nodes between different independent views through cosine similarity, and taking the similarity as second data; projecting the potential features into a query or key space, and calculating an attention weight between two nodes of different independent views as third data; constructing a soft reconstruction subgraph matrix based on the second data and the third data, and obtaining a corresponding feature matrix; forming a fusion feature matrix based on the potential features and features corresponding to the soft reconstruction sub-graphs; processing the fusion feature matrix through a multi-layer linear function to obtain a fractional matrix corresponding to the category of each node; a total loss is calculated based on the fractional matrix.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent transportation systems, and more particularly relates to an uncertainty-aware cross-graph learning method for conflict data truth inference in vehicular crowdsourcing. Background Art

[0002] Vehicular Crowdsensing (VCS) uses ubiquitous in-vehicle sensors to collect data and shows great potential in achieving low-cost, large-scale sensing tasks. However, due to the differences in the professionalism of these vehicles, they may provide conflicting information. Therefore, a fundamental challenge in current VCS applications is how to infer accurate answers from conflicting data, which has prompted the emergence of truth inference research.

[0003] Previous studies have attempted to determine the accuracy of users' answers by evaluating the quality of service of sensing users. This method usually requires evaluating each task of each worker, resulting in many unnecessary costs (e.g., time and money). In addition, due to the subjectivity of evaluation metrics, even if these workers have high service quality, they may still output incorrect results. Some studies have addressed the above problems by mining the potential relationships in VCS, such as the correlation between workers or the relationship between workers and tasks. However, since they need to model for specific sensing tasks, all these solutions are only applicable to limited scenarios / tasks.

[0004] Therefore, designing a better generalization model has become one of the urgent problems to be solved in truth inference. In recent years, many scholars have used graph neural networks to convert each data source in VCS into a graph, where data points are used as nodes and edges represent the relationships between data points. Subsequently, they infer accurate answers by fusing the graphs corresponding to multiple data sources. Unfortunately, although the above solutions have achieved breakthrough results in truth inference, in order to enhance the applicability of graph-based truth inference methods in more practical VCS applications, two core scientific problems must be solved:

[0005] How to effectively capture the consistency and complementarity information among crowdsourced data. The heterogeneity of VCS data is manifested between different nodes in the same view and also between multiple views. For the internal of a view, most existing graph-based ground truth inference methods cannot consider the different contributions of different nodes in information aggregation. It is difficult to accurately describe the underlying information of each data source by equally aggregating node features. In addition, current fusion methods are generally divided into two categories. The first type of method aims to maximize the similarity between views and usually adopts a hard clustering mechanism (such as K-means) to set a fixed threshold to eliminate noisy data. This method ignores that complementarity is another important feature of multi-source data in VCS. Since cross-view node collaboration can provide more comprehensive target information, filtering based only on similarity may discard valuable but dissimilar data, resulting in potential information loss. Another common method is to assign different weights to different views and fuse them into a consensus graph to predict the true labels of the perceived targets. However, it ignores the heterogeneity of the information content and topology across views, which dynamically evolve with the change of tasks. This defect may hinder the preservation of key topological relationships and complementary information between views, ultimately limiting the performance of model inference. The above problems indicate that there is an urgent need for innovative information integration methods across multiple data sources in VCS to ensure the effective fusion of information within the same view and between different views simultaneously.

[0006] How to evaluate the credibility of model output. Most ground truth inference strategies focus on training models to improve prediction accuracy while ignoring the confidence of model decision results. The inherent uncertainty in model responses may pose significant risks to the safety of VCS applications. For example, when applying a ground truth inference strategy to object recognition in autonomous vehicles, the model may have difficulty determining the best choice for an object with multiple labels and similar prediction probabilities. This ambiguity may lead to serious errors and even traffic accidents. Therefore, a new method must be developed to evaluate the confidence of model decisions in the context of VCS ground truth inference to ensure accuracy and reliability in practical applications. Summary of the Invention

[0007] The present invention is precisely proposed based on the above-mentioned requirements of the prior art. The technical problem to be solved by the present invention is to provide an uncertainty-aware cross-graph learning method for ground truth inference of conflicting data in vehicle crowdsourcing to effectively capture the consistency and complementarity information among crowdsourced data and simultaneously ensure the accuracy and reliability of confidence.

[0008] To solve the above problems, the technical solutions provided by the present invention include:

[0009] An uncertainty-aware cross-graph learning method for true value inference of conflicting data in vehicle crowdsourcing is provided, including: obtaining multiple independent views formed by multiple different observation perspectives corresponding to the perception tasks issued by the data requester, where the independent views include a node set, an edge set, and a node feature matrix, and the node set includes multiple nodes; calculating the attention coefficients between the nodes in the node set within the independent views as the first data, and stacking graph attention layers to obtain latent features; calculating the similarity between pairwise nodes of different independent views based on the latent features through cosine similarity as the second data; projecting the latent features into the query or key space, and calculating the normalized attention weights between the two nodes of different independent views as the third data; calculating and constructing a soft reconstruction subgraph matrix based on the second data and the third data; constructing a corresponding feature matrix based on the soft reconstruction subgraph matrix; obtaining fusion features based on the latent features and the features in the feature matrix corresponding to the soft reconstruction subgraph, and forming a fusion feature matrix; processing the fusion feature matrix through a multi-layer linear function to obtain a score matrix corresponding to the category of each node; calculating the total loss based on the score matrix, where the total loss is obtained by weighting the true value loss and the Bayesian loss.

[0010] Preferably, the first data is expressed as: is the attention coefficient between node i and node j in the t-th view, that is, the first data, representing the importance of node j to node i. represents the attention mechanism. is the feature of the i-th node. is the feature of the j-th node.

[0011] Preferably, the latent feature is expressed as: Among them, is the i-th node in the t-th perspective, σ() represents the non-linear activation function, l represents the number of layers, i is the perspective serial number, and t is the node serial number.

[0012] Preferably, the second data is expressed as: Among them, · represents the dot product, ||·|| represents the Euclidean norm of the vector, S ij (u,v) represents the similarity between the i-th node in the u-th view and the j-th node in the v-th view, represents the latent feature of the i-th node in the u-th view, represents the latent feature of the j-th node in the v-th view.

[0013] Preferably, the third data is expressed as: Among them, W q , W k ∈R d ×dare learnable parameters, d is the feature dimension, is the normalized attention weight between two nodes in views u and v, that is, the third data.

[0014] Preferably, the soft reconstruction subgraph matrix S (u) ∈R N×N , and each element is calculated as:

[0015] is the weighted similarity representation of the relationship between nodes i and j in subgraph u.

[0016] Preferably, the feature matrix corresponding to the u-th subgraph is expressed as: In the formula, represents the reconstructed subgraph S (u) is the feature matrix of all nodes included after being processed by the feature-enhanced graph attention encoder, is a transformation matrix, where D sub is the dimension of the subgraph features, and D is the dimension of the features.

[0017] Preferably, the fusion feature matrix its elements are expressed as: where is the fusion feature of node j.

[0018] Preferably, when each sample predicts a single class, the sample label is predicted through the Softmax function, and y i is the predicted true label of sample i: y i,k = argmax k z prob(i,k) , and the corresponding true value loss is: where Z i,j is the vector in the score matrix, representing the probability that the i-th sample belongs to the k-th class; when each sample belongs to multiple classes, the scores of each class are calculated separately, and the sample class probabilities are transformed using the Sigmoid function, θ is the threshold. For the multi-label true value inference problem, the true value loss is calculated as follows:

[0019] Preferably, the Bayesian loss calculation in the single-label classification scenario: e i,k = 1 - α i,k , and the Bayesian loss calculation in the multi-label classification scenario: U i,k = (y i,k - P b(i,k) )2 , The total loss is: L = ψL truth + τL bayes where ψ and τ are both learnable weight parameters that can be adjusted during training to optimize the model performance.

[0020] Compared with the prior art, in order to process incomplete information captured from multiple data sources, we propose a hierarchical fine-grained fusion strategy. This scheme not only includes information integration among nodes within a specific view but also realizes cross-view information propagation. To address the problem that the model output lacks confidence evaluation, based on the principle of evidential deep learning, a new method for evaluating the confidence of the true value inference model is proposed to ensure that the results output by the model are true. Based on the above invention content, two loss functions are designed in the model optimization strategy to jointly optimize the UCGL algorithm and realize the autonomous learning iteration of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] To more clearly illustrate the technical solutions in the embodiments of this specification or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments recorded in the embodiments of this specification. For those of ordinary skill in the art, other drawings can also be obtained based on these drawings.

[0022] Figure 1 It is a flowchart of the steps of an uncertainty-aware cross-graph learning method for true value inference of conflict data in vehicle crowdsourcing in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0024] In the description of the embodiments of the present invention, it should be noted that unless otherwise clearly defined and limited, the term "connected" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection. It can be a mechanical connection or an electrical connection. It can be directly connected or indirectly connected through an intermediate medium. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0025] The terms "top", "bottom", "above", "below", and "on" used throughout the description are relative positions with respect to the components of the device, such as the relative positions of the top and bottom substrates inside the device. It is understood that the devices are multifunctional and independent of their orientation in space.

[0026] To facilitate the understanding of the embodiments of the present invention, the following will further explain and illustrate with specific embodiments in conjunction with the accompanying drawings. The embodiments do not constitute a limitation to the embodiments of the present invention.

[0027] This embodiment provides an uncertainty-aware cross-graph learning method for inferring the ground truth of conflicting data in vehicle crowdsourcing, as Figure 1 shown.

[0028] The uncertainty-aware cross-graph learning method for inferring the ground truth of conflicting data in vehicle crowdsourcing includes:

[0029] Obtaining independent views formed by multiple different observation perspectives corresponding to the perception tasks issued by the data requester, where the independent views include a node set, an edge set, and a node feature matrix.

[0030] The VCS system usually consists of a data requester, some perception participants, and a cloud platform. The data requester directly sends the specific requirements of the perception task, such as the task object, the location of the task, the type of uploaded data, etc., to the cloud platform, and requests the ground truth label of the perception task. Subsequently, the cloud platform releases the task to the interested vehicles according to the time and location of the perception task. After receiving the task, the perception vehicles return the collected data to the cloud platform.

[0031] For the perception tasks issued by the data requester, after the perception vehicles receive the tasks, multiple different observation perspectives are formed. The data obtained from each perspective is regarded as forming a separate view. A separate view includes a node set, an edge set indicating whether there is an association between nodes, and a node feature matrix. Nodes represent data samples, and edges represent the association relationships between samples.

[0032] Specifically, the separate view is denoted as G t ={V t ,E t ,H t}, t ∈ m. Where V t ={v i} i＝1,...n represents the node set, represents the nth sample (node) in the tth view. E = {e ij} represents the edge set indicating whether there is an association between node i and node j, represents the node feature matrix of the tth view, F represents the number of features of each node. And Denote the attributes / features of the $i$-th node in the $t$-th view. $m$ is the number of acquired observation perspectives.

[0033] Capture consistent and complementary information from the above multi-view data, and classify each node into $K$ unrelated categories. Depending on the task, the true label of a sample can have one or more labels. Finally, report the true value inference result to the data requester.

[0034] Calculate the attention coefficients between nodes in independent views as the first data, and stack the graph attention layer to obtain latent features.

[0035] Apply the weight matrix $W\in\mathbb{R}$ to each node F×F , and at the same time combine the attention mechanism, and calculate the attention coefficients between nodes to characterize the importance of other nodes adjacent to a certain node to this node, specifically expressed as:

[0036]

[0037] $\alpha_{ij}^t$ is the attention coefficient between node $i$ and node $j$ in the $t$-th view, that is, the first data, indicating the importance of node $j$ to node $i$. Denote the attention mechanism $h_i^t$ is the feature of the $i$-th node $h_j^t$ is the feature of the $j$-th node.

[0038] To compare the importance of different nodes $j$ to node $i$, normalize the first data through the softmax function, expressed as:

[0039] Based on the normalized first data, stack the graph attention layer to obtain the latent feature of each node $i$.

[0040] Specifically expressed as where $\sigma()$ represents the non-linear activation function, $l$ represents the number of layers, $i$ is the view serial number, and $t$ is the sample serial number.

[0041] When $l = 1$, there is

[0042] When $l = 2$, there is

[0043] Assign to That is, the obtained latent feature is expressed as

[0044] Based on the latent features, calculate the similarity between pairwise nodes in different independent views through cosine similarity as the second data.

[0045] The specific calculation is as follows:

[0046]

[0047] Where, · represents the dot product, and ||·|| represents the Euclidean norm of the vector. S ij (u, v) represents the similarity between the i-th node of the u-th view and the j-th node of the v-th view. represents the latent feature of the i-th node of the u-th view, represents the latent feature of the j-th node of the v-th view.

[0048] Project the latent features into the query or key space, and calculate the normalized attention weights between two nodes of different independent views as the third data.

[0049]

[0050] Where, W q , W k ∈R d×d are learnable parameters, d is the feature dimension, is the normalized attention weight between two nodes in views u and v, that is, the third data.

[0051] Calculate and construct the soft reconstruction subgraph matrix based on the second data and the third data.

[0052] Use attention-weighted similarity to dynamically aggregate edges from multiple views, and define the soft reconstruction subgraph matrix S (u) ∈R N×N , where each element is calculated as:

[0053]

[0054] is the weighted similarity representation of the relationship between nodes i and j in subgraph u, which is used to reconstruct the adjacency matrix, so that the structure of each subgraph can be dynamically adjusted according to different data sources instead of being fixed.

[0055] Construct the corresponding feature matrix based on the soft reconstruction subgraph matrix.

[0056] Calculate the feature matrix corresponding to the u-th subgraph

[0057]

[0058] In the formula, represents the feature matrix of all nodes contained in the reconstructed subgraph S (u) after being processed by the feature-enhanced graph attention encoder. is a transformation matrix that ensures that the feature matrix of the reconstructed subgraph after being processed by the feature-enhanced graph attention encoder is consistent with the dimension of the features. Among them, D sub is the dimension of the subgraph features, and D is the dimension of the features of

[0059] This step ensures that similar nodes are grouped together to generate a new adjacency matrix. The key motivation behind this new method of reshaping the graph structure is that traditional hard clustering methods often suffer from the problem of rigid boundaries. For example, setting a fixed similarity threshold to determine whether there is an edge between two points can easily lead to information loss. In addition, there may be a problem of confusion in subgraph descriptions because there are nodes that may contain multiple attributes related to various categories. Therefore, the method we proposed aims to enable nodes to participate in multiple subgraphs with different degrees of relevance, thus allowing for more flexible and efficient representation learning. To solve the subgraph

[0060] Based on the latent features and the features corresponding to the soft reconstructed subgraph, the fused features are obtained and a fused feature matrix is formed.

[0061] Specifically expressed as:

[0062]

[0063] Among them, is the fused feature of node j.

[0064] Based on the feature-enhanced graph attention encoder, the processed node features are obtained Suppose the j-th node in the t-th view (hereinafter referred to as node j) is divided into s reconstructed subgraphs. For each reconstructed subgraph u, the feature of node j is defined as Integrate the node feature information of all reconstructed subgraphs covering node j and

[0065] Integrate the fused features of all the above feature information to form a fused feature matrix and where N represents the number of samples and D represents the feature dimension.

[0066] The fused feature matrix is processed by a multi-layer linear function to obtain the score matrix corresponding to the category of each node.

[0067] Specifically, the multi-layer linear function obtains the score matrix Z = Z corresponding to the category K of each sample (l) ∈R N×K .

[0068]

[0069] Z(L) = σ(Z (L-1) W (L) + b (L) )

[0070] L = 1, 2...l

[0071] is the weight matrix of layer L, b (L) ∈ R 1＝K is the bias term of layer L.

[0072] Calculate the true value loss based on the score matrix

[0073] True value loss: The true value loss is used to measure the probability that the model predicts the true label. True value loss functions will be constructed for single-label true value inference and multi-label true value inference cases respectively.

[0074] If each sample predicts a single class, then the sample label is predicted through the Softmax function, y i is the predicted true value label of sample i:

[0075]

[0076] The corresponding true value loss is:

[0077]

[0078] where, Z i,j is the vector in the score matrix, representing the probability that the i-th sample belongs to the k-th class.

[0079] If each sample can belong to multiple classes. Calculate the scores for each class respectively, and transform the sample class probabilities using the Sigmoid function.

[0080]

[0081] θ is the threshold and can be adjusted according to the actual situation.

[0082] For the multi-label true value inference problem, the true value loss is calculated as follows:

[0083]

[0084] Calculate the Bayesian loss based on the score matrix.

[0085] For the inference problem of K classes, each class k in sample i consists of b i,k and u i,k which respectively represent the model's belief in each class k in sample i and the overall uncertainty related to the prediction,

[0086]

[0087] e i,k = 1 - α i,k

[0088] α i,k is a parameter of the Dirichlet distribution.

[0089] In this embodiment, e i,k = Softplus(Z i,k ) is used to replace the above formula to obtain the uncertainty of the model, and thus the confidence of the model answer. In this model, we use the activation layer Softplus instead of the Softmax layer to ensure non - negative output as the evidence vector for predicting the Dirichlet distribution.

[0090] Bayesian loss: This patent uses Bayesian loss to quantify the confidence of the model. Using the following formula, Bayesian loss functions are designed for single - label true - value inference problems and multi - label true - value inference problems. Among them, P b represents the posterior probability. U is the square of the deviation, which reflects the deviation between the confidence of the model in the correct class and the actual situation. V represents the variance, indicating the confidence and uncertainty of this class in the model.

[0091] Calculation of Bayesian loss in the single - label classification scenario:

[0092]

[0093] Calculation of Bayesian loss in the multi - label classification scenario:

[0094] U i,k = (y i,k - P b(i,k) ) 2

[0095]

[0096] Combine the true - value loss and the Bayesian loss to calculate the total loss.

[0097] We unify the true - value loss and the Bayesian loss and jointly optimize the above - mentioned scheme, where the total loss is:

[0098] L = ψL truth + τL bayes

[0099] Among them, both ψ and τ are learnable weight parameters that can be adjusted during training to optimize the model performance.

[0100] In the above steps, we use the designed feature-enhanced graph attention encoder to enhance the features of each node in the inner view. Subsequently, based on the feature similarity of nodes across views, we adopt the proposed soft subgraph reconstruction strategy to reorganize all nodes in multiple data sources into multiple subgraphs. However, since some nodes may contain multiple potential category feature information, they may be divided into different subgraphs. To provide effective supplementary information for downstream tasks, we integrate the feature information of the nodes enhanced in the feature-enhanced encoder with the feature information of the node across different reconstructed subgraphs.

[0101] The above work only improves the accuracy of the ground truth inference results, but does not provide the credibility of the answers. In addition, previous ground truth inference models judge the confidence of the ground truth inference results by evaluating the credibility of the participants. These credibility evaluations are more subjective and do not directly judge the credibility of the model answers. In some cases, even if the participants are honest, the staff is not sure about the answers because the probability distributions of the inference results are very close. Therefore, we need to provide a more robust method to judge the confidence of the inference results.

[0102] Evidential Deep Learning (EDL) is a confidence evaluation method based on Dempster-Shafer evidence theory. It provides a more objective and data-driven method to evaluate model uncertainty, thus overcoming the limitations of the above traditional methods. Currently, EDL has been widely applied to the confidence research of various inference models. EDL attempts to directly learn the parameters of the distribution through deep learning to understand and quantify model uncertainty. In EDL, the uncertainty of model predictions is modeled as a distribution of probabilities through the Dirichlet distribution.

[0103] In the above steps, we use the designed feature-enhanced graph attention encoder to enhance the features of each node in the inner view. Subsequently, based on the feature similarity of nodes across views, we adopt the proposed soft subgraph reconstruction strategy to reorganize all nodes in multiple data sources into multiple subgraphs. However, since some nodes may contain multiple potential category feature information, they may be divided into different subgraphs. To provide effective supplementary information for downstream tasks, we integrate the feature information of the nodes enhanced in the feature-enhanced encoder with the feature information of the node across different reconstructed subgraphs.

[0104] Based on the feature-enhanced graph attention encoder, we obtain the processed node features Suppose the j-th node in the t-th view (hereinafter referred to as node j) is divided into s reconstructed subgraphs. For each reconstructed subgraph u, we define the feature of node j as This paper integrates the node feature information of all reconstructed subgraphs covering node j with

[0105]

[0106] Represents the final feature representation of node j.

[0107] The specific embodiments described above further elaborate on the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only the specific embodiments of the present invention and is not used to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. An uncertainty-aware cross-graph learning method for true value inference of conflicting data in vehicle crowdsourcing, characterized in that Including: Obtain multiple independent views formed by multiple different observation perspectives corresponding to the perception tasks issued by the data requester. The independent views include a node set, an edge set, and a node feature matrix, and the node set includes multiple nodes; Calculate the attention coefficients between the nodes in the node set within the independent view as the first data, and stack the graph attention layer to obtain latent features; Based on the latent features, calculate the similarity between pairwise nodes of different independent views through cosine similarity as the second data; Project the latent features into the query or key space, and calculate the normalized attention weights between the two nodes of different independent views as the third data; Calculate and construct a soft reconstruction subgraph matrix based on the second data and the third data; Construct a corresponding feature matrix based on the soft reconstruction subgraph matrix; Obtain fused features based on the latent features and the features in the feature matrix corresponding to the soft reconstruction subgraph, and form a fused feature matrix; Process the fused feature matrix through a multi-layer linear function to obtain a score matrix corresponding to the category of each node; Calculate the total loss based on the score matrix, and the total loss is obtained by weighting the ground-truth loss and the Bayesian loss.

2. The uncertainty-aware cross-graph learning method for true value inference of conflicting data in vehicle crowdsourcing according to claim 1, characterized in that The first data is expressed as: is the attention coefficient between node i and node j in the t-th view, that is, the first data, representing the importance of node j to node i. represents the attention mechanism. is the feature of the i-th node. is the feature of the j-th node.

3. The uncertainty-aware cross-graph learning method for true value inference of conflicting data in vehicle crowdsourcing according to claim 2, characterized in that The latent features are expressed as: where, is the i-th node of the t-th perspective, σ( ) represents the non-linear activation function, l represents the number of layers, i is the perspective number, and t is the node number.

4. The uncertainty-aware cross-graph learning method for true value inference of conflicting data in vehicle crowdsourcing according to claim 3, characterized in that The second data is expressed as: where · represents the dot product, ||·|| represents the Euclidean norm of a vector, and S ij (u, v) represents the similarity between the i-th node of the u-th view and the j-th node of the v-th view, represents the latent feature of the i-th node of the u-th view, represents the latent feature of the j-th node of the v-th view.

5. The uncertainty-aware cross-graph learning method for true value inference of conflicting data in vehicle crowdsourcing according to claim 4, characterized in that The third data is expressed as: Among them, W q , W k ∈R d×d are learnable parameters, d is the feature dimension, is the normalized attention weight between two nodes in view u and view v, that is, the third data.

6. The uncertainty-aware cross-graph learning method for true value inference of conflicting data in vehicle crowdsourcing according to claim 5, characterized in that Soft reconstructed subgraph matrix S (u) ∈R N×N , where each element is calculated as: It is the weighted similarity representation of the relationship between node i and node j in subgraph u.

7. The uncertainty-aware cross-graph learning method for true value inference of conflicting data in vehicle crowdsourcing according to claim 6, characterized in that The feature matrix corresponding to the u-th subfigure It is expressed as: In the formula, represents the reconstructed subgraph S (u) is the feature matrix of all nodes included after being processed by the feature-enhanced graph attention encoder, is a transformation matrix, where D sub is the dimension of the subgraph features, and D is the dimension of the features.

8. The uncertainty-aware cross-graph learning method for true value inference of conflicting data in vehicle crowdsourcing according to claim 7, characterized in that The fused feature matrix The elements thereof are represented as: Among them, is the fused feature of node j.

9. The uncertainty-aware cross-graph learning method for true value inference of conflicting data in vehicle crowdsourcing according to claim 8, characterized in that When each sample predicts a single class, the sample label is predicted by the Softmax function, y i is the true label predicted for sample i: The corresponding ground-truth loss is: Among them, Z i,j is a vector in the fractional matrix, representing the probability that the i-th sample belongs to the k-th category; When each sample belongs to multiple categories, calculate the scores of each category respectively, and transform the sample category probability with the Sigmoid function, θ is the threshold. For the multi-label ground-truth inference problem, the ground-truth loss is calculated as follows:

10. The uncertainty-aware cross-graph learning method for true value inference of conflicting data in vehicle crowdsourcing according to claim 9, characterized in that Calculation of Bayesian loss in single-label classification scenario: e i,k ＝1-α i,k Calculation of Bayesian loss in multi-label classification scenario: U i,k = (y i,k - P b(i,k) ) 2 The total loss is: L = ψL truth + τL bayes Where ψ and τ are both learnable weight parameters that can be adjusted during training to optimize the model performance.

Images