A Heterogeneity-Guided Traffic Flow Completion Method, System, Medium, and Device

Abstract

The present application provides a method, system, medium, and device for heterogeneous-guided traffic flow completion, belonging to the technical field of data processing. The method includes: performing spatio-temporal decoupled masked encoding on the original traffic flow data to generate masked data; respectively performing feature embedding and positional encoding on the original traffic flow data and the masked data to obtain spatio-temporal encoding results; wherein the spatio-temporal encoding results include the temporal and spatial encoding results of the original traffic flow data and the masked data respectively; based on the spatio-temporal encoding results, using a masked contrastive learning method of reconstruction loss, respectively generating a temporal heterogeneity representation matrix and a spatial heterogeneity representation matrix; using the temporal heterogeneity representation matrix and the spatial heterogeneity representation matrix as constraint conditions, and completing the traffic flow data based on a tensor decomposition model to obtain the completed traffic flow. By using the spatio-temporal heterogeneity representation results to constrain the spatio-temporal relationship of the traffic flow, the completion accuracy and robustness of the road traffic flow data are improved.

Description

Technical Field

[0001] This application relates to the technical field of data processing, and particularly to a heterogeneity-guided traffic flow completion method, system, medium, and device. Background Art

[0002] Accurate and reliable traffic flow data is particularly important for the research and application of intelligent transportation systems, and is the basic data guarantee for exerting its precise regulation and decision-making support functions. Currently, the acquisition of road traffic flow information mainly relies on professional fixed detection devices (such as cameras, radar / laser speed detectors, etc.). Restricted by the deployment density and communication transmission conditions, there are inevitably sparse or missing phenomena in the traffic flow records collected by fixed detection points on the roads, which restricts the acquisition of accurate and reliable traffic flow data. Therefore, how to complete the missing data from partial observation data of traffic flow monitoring stations has been widely concerned in the industry.

[0003] The completion of road traffic flow data refers to inferring the missing flow data using techniques such as interpolation, statistical methods, and machine learning. As an important branch of machine learning, tensor decomposition methods have been widely used in traffic flow data completion due to their advantages in high-dimensional data processing. However, this method has biases in capturing the spatio-temporal distribution patterns of traffic flow, especially in sparse or non-stationary environments, affecting the accuracy and robustness of completion.

[0004] Therefore, an improved technical solution is needed to address the deficiencies of the above-mentioned existing technologies. Summary of the Invention

[0005] The purpose of this application is to provide a heterogeneity-guided traffic flow completion method, system, medium, and device to solve or alleviate the problems existing in the above-mentioned existing technologies.

[0006] To achieve the above purpose, this application provides the following technical solutions:

[0007] In the first aspect, this application provides a heterogeneity-guided traffic flow completion method, including the following steps:

[0008] Perform spatio-temporal decoupled mask encoding on the original traffic flow data to generate mask data , where the mask data includes temporal mask data and spatial mask data ; wherein, the original traffic flow data is traffic flow data with missing values, which is represented in tensor form to obtain the original traffic flow tensor ;

[0009] Perform feature embedding and positional encoding on the original traffic flow data and the mask data respectively to obtain a spatio-temporal encoding result; wherein, the spatio-temporal encoding result includes the time and space encoding results of the original traffic flow data and the mask data respectively;

[0010] Based on the spatio-temporal encoding result, use the mask contrast learning method of reconstruction loss to generate a time heterogeneity representation matrix and a space heterogeneity representation matrix respectively;

[0011] Take the time heterogeneity representation matrix and the space heterogeneity representation matrix as constraint conditions, and complete the traffic flow data based on the tensor decomposition model to obtain the completed traffic flow 。

[0012] Preferably, the generation method of the mask data is as follows:

[0013] According to the original traffic flow tensor , along the time dimension and the space dimension respectively according to the masking rate Randomly generate a time mask tensor And a space mask tensor ;

[0014] Multiply the original traffic flow tensor With the time mask tensor , the space mask tensor Perform an element-wise multiplication operation respectively to obtain the mask data 。

[0015] Preferably, perform feature embedding and positional encoding on the original traffic flow data and the mask data respectively to obtain a spatio-temporal encoding result, specifically:

[0016] Adopt a combination of block embedding technology and a fully connected layer to generate high-dimensional embedding features of the original traffic flow data and the mask data respectively;

[0017] Use the sine positional encoding technology to generate the spatio-temporal positional encodings of the original traffic flow data and the mask data respectively;

[0018] Multiply the high-dimensional embedding features of the time dimension and the space dimension of the original traffic flow data and the mask data respectively with the spatio-temporal positional encodings of the original traffic flow data and the mask data respectively, and then input them into the Transformer to obtain the spatio-temporal encoding result.

[0019] Preferably, based on the spatio-temporal encoding result, use the mask contrast learning method of reconstruction loss to generate a time heterogeneity representation matrix and a space heterogeneity representation matrix respectively, including:

[0020] Perform mean and dimensionality transformation operations on the time encoding results of the original traffic flow data and the mask data respectively to extract the average feature representations of each time step of the original traffic flow data and the mask data. and ;

[0021] Introduce the first parameter matrix and the second parameter matrix , and multiply the first parameter matrix , the second parameter matrix respectively with the average feature representations and of each time dimension of the original traffic flow data and the mask data to obtain the time heterogeneity characterization results and ;

[0022] Calculate and to obtain the first loss ;

[0023] Perform mean operations on the spatial encoding results of the original traffic flow data and the mask data respectively in the embedding dimension to extract the average feature representations of the original traffic flow data and the mask data at each node position and ;

[0024] Introduce the third parameter matrix and the fourth parameter matrix , and multiply the third parameter matrix , the fourth parameter matrix respectively with the average feature representations of the original traffic flow data and the mask data at each node position and to obtain the spatial heterogeneity characterization results and ;

[0025] Calculate and to obtain the second loss ;

[0026] Generate the final time heterogeneity characterization matrix and and the spatial heterogeneity characterization matrix and and by iteratively optimizing the first loss and the second loss .

[0027] Preferably, the tensor decomposition model is a CP decomposition model, and the original traffic flow tensor is decomposed into a first factor matrix by using the CP decomposition model , a second factor matrix , and a third factor matrix , and its probability distribution is optimized by variational Bayesian learning.

[0028] Preferably, the loss function of the CP decomposition model is the sum of the following four parts: a reconstruction error term, a spatial heterogeneity representation constraint term, a temporal heterogeneity representation constraint term, and a regularization term;

[0029] The reconstruction error term is the reconstruction error between the original traffic flow and the completed traffic flow;

[0030] The spatial heterogeneity representation constraint term is used to constrain the spatial relationship of traffic flows between different nodes, and its expression is:

[0031] ,

[0032] In the formula, is the spatial heterogeneity representation matrix of the original traffic flow data; is the first factor matrix;

[0033] The temporal heterogeneity representation constraint term is used to constrain the temporal relationship of traffic flows between different nodes, and its expression is as follows:

[0034] ,

[0035] In the formula, is the temporal heterogeneity representation matrix of the original traffic flow data, , are the elements of the first factor matrix indexed by ( ), ( ), is the element of the second factor matrix indexed by ( );

[0036] The expression of the regularization term is: .

[0037] Preferably, the weights of the first factor matrix , the second factor matrix , and the third factor matrix are initialized by using the reparameterization trick, specifically:

[0038] ,

[0039] ,

[0040] ,

[0041] In the formula, , , are the weights of the first factor matrix , the second factor matrix , and the third factor matrix respectively. , , are the global parameters corresponding to the first factor matrix , the second factor matrix , and the third factor matrix respectively. is the standard deviation of the variational posterior parameter . Conforms to a Gaussian distribution. , , are the random noises corresponding to the first factor matrix , the second factor matrix , and the third factor matrix respectively. , , All follow a normal distribution.

[0042] In a second aspect, this embodiment provides a heterogeneity-guided traffic flow completion system, including:

[0043] A mask encoding module for performing spatio-temporal decoupled mask encoding on the original traffic flow data to generate mask data , where the mask data includes temporal mask data and spatial mask data ; wherein, the original traffic flow data is traffic flow data with missing values, which is represented in tensor form to obtain the original traffic flow tensor ;

[0044] An embedding and position encoding module for respectively performing feature embedding and position encoding on the original traffic flow data and the mask data to obtain spatio-temporal encoding results; wherein, the spatio-temporal encoding results include the temporal and spatial encoding results of the original traffic flow data and the mask data respectively.

[0045] A spatio-temporal heterogeneity representation module, configured to generate a temporal heterogeneity representation matrix and a spatial heterogeneity representation matrix respectively based on the spatio-temporal encoding result by using a masked contrastive learning method of reconstruction loss;

[0046] A decomposition module, configured to use the temporal heterogeneity representation matrix and the spatial heterogeneity representation matrix as constraint conditions to complete traffic flow data based on a tensor decomposition model, and obtain the completed traffic flow .

[0047] In a third aspect, an embodiment of the present application provides an electronic device, including: a memory, configured to store instructions executed by one or more processors of the electronic device; a processor, when the processor executes the instructions in the memory, enabling the electronic device to implement the steps of the heterogeneity-guided traffic flow completion method described in any of the above embodiments.

[0048] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, characterized in that instructions are stored on the computer-readable storage medium, and when the instructions are executed on a computer, the steps of the heterogeneity-guided traffic flow completion method described in any of the above embodiments are implemented.

[0049] The technical solution of the embodiment of the present application has the following beneficial effects:

[0050] By performing spatio-temporal decoupled masked encoding on the original traffic flow data to obtain temporal masked data and spatial masked data, it is possible to capture the spatio-temporal change characteristics of traffic flow with rich context, and introduce masked contrastive learning based on reconstruction loss in the process of spatio-temporal heterogeneity representation to enhance the model's ability to reconstruct the spatio-temporal heterogeneity representation of missing positions; using the spatio-temporal heterogeneity representation results (i.e., the temporal heterogeneity representation matrix and the spatial heterogeneity representation matrix) to constrain the spatio-temporal relationship of traffic flow, and optimizing the spatio-temporal representation process and the traffic flow completion process through self-supervised learning to enhance the completion performance of road traffic flow data. In summary, the method provided in this embodiment can effectively represent the spatio-temporal heterogeneity of road traffic flow data under different missing patterns, and exhibit high traffic flow data completion accuracy and robustness. Description of the Drawings

[0051] Figure 1 It is a schematic flowchart of a heterogeneity-guided traffic flow completion method provided according to some embodiments of the present application.

[0052] Figure 2 It is a logical block diagram of a heterogeneity-guided traffic flow completion method provided according to some embodiments of the present application.

[0053] Figure 3 It is a schematic structural diagram of an electronic device provided according to some embodiments of the present application. Detailed implementation manners

[0054] To facilitate the understanding of the technical solution of this application, the following will make a detailed description of relevant terms and the proposed ideas of the present invention.

[0055] The tensor decomposition method is a mathematical tool for processing high-order data (i.e., multi-dimensional arrays). It decomposes a high-dimensional tensor into a combination of multiple low-dimensional structures to extract latent features, reduce dimensions or denoise.

[0056] Common tensor decomposition methods include the CP model, the Tucker model, the Bayesian decomposition model, and the rank minimization model. Such methods usually utilize context features such as points of interest (POIs) and road networks around monitoring points to model the relationship of traffic flows between different regions, and then constrain the tensor decomposition process to improve the accuracy of traffic flow completion.

[0057] However, traffic flow is non-linearly affected by various factors such as geographical environment, infrastructure, and travel demand, showing obvious spatio-temporal heterogeneity characteristics, that is, traffic flow has significant differences at different spatial positions and time periods. For example, the traffic flow in the business district may be higher on weekends, while the opposite is true in the residential area; there is an obvious morning rush hour on certain roads on weekdays, but the traffic flow is more uniform on weekends. Existing tensor decomposition methods usually rely on the "spatial similarity assumption", that is, it is assumed that regions with similar geographical features have similar traffic flow change trends, ignoring the impact brought by spatio-temporal heterogeneity. This assumption leads to biases in the model when capturing the spatio-temporal distribution pattern of traffic flow, especially in sparse or non-stationary environments, affecting the accuracy and robustness of completion.

[0058] In recent years, the introduction of spatio-temporal heterogeneity characterization methods has significantly promoted the development of road traffic data modeling techniques. According to the principles of this method, it is mainly divided into two categories: meta-learning and representation learning. Among them, meta-learning methods rely on context features such as points of interest (POIs), monitoring point locations, or GPS trajectories, aiming to obtain parameters that can effectively represent the characteristics of different regions, thereby enhancing the model's ability to represent spatio-temporal heterogeneity. However, the effectiveness of these models may be limited by the availability of context features. The representation learning method aims to start from a data-driven perspective and use feature embedding or spatio-temporal encoding to mine the implicit spatio-temporal heterogeneity in traffic data, providing a new reference for traffic volume data completion based on tensor decomposition. Among them, self-supervised learning has become one of the key strategies for enhancing the generalization ability and adaptability of spatio-temporal heterogeneity representation due to its effectiveness in enhancing the representation of spatio-temporal data features. This type of method helps the model capture complex spatial and temporal dependencies in spatio-temporal data by designing appropriate pre-training tasks, thereby enhancing the model's understanding and representation ability of data, and further promoting the accuracy and effectiveness of spatio-temporal heterogeneity representation. Therefore, combining self-supervised learning technology with tensor decomposition is an ideal solution, which is expected to improve the accuracy and adaptability of completion. However, the inventors have found that the above combination faces two major challenges in practice:

[0059] (1) Existing self-supervised learning methods are usually established based on complete historical data, allowing the model to find spatio-temporal patterns and dependencies in the entire spatio-temporal range. However, different degrees of missing traffic flow data may disrupt its continuity and consistency, resulting in inaccurate mining of spatio-temporal distribution patterns, thereby affecting the representation effect of spatio-temporal heterogeneity. For example, the traffic flow changes during peak hours on certain sections may be very different from those during non-peak hours, and missing data may mask this change. Therefore, when combining self-supervised learning technology with tensor decomposition, the model often has difficulty effectively identifying and representing the spatio-temporal heterogeneity features of missing positions, resulting in the inability to optimize according to the true data distribution.

[0060] (2) Due to the lack of a supervision signal for missing positions in the optimization process of the tensor decomposition method, it is impossible to reverse-optimize the spatio-temporal heterogeneity representation based on the completed traffic flow results. The absence of this feedback mechanism interrupts the mutual promotion between representation learning and data completion, resulting in the model's inability to adaptively adjust the representation of spatio-temporal heterogeneity features according to the completed results, thereby limiting the model's adaptability to different missing patterns and complex traffic flow change trends. Especially in data-sparse and non-stationary environments, the structure lacking effective feedback significantly weakens the model's ability to capture heterogeneity features and improve generalization performance.

[0061] In view of this, the present application proposes a heterogeneity-guided traffic flow completion method (also known as: heterogeneity-guided tensor decomposition method, abbreviated as: Hg-TD), which can effectively perform spatio-temporal heterogeneity characterization in the case of missing traffic flow data and lack of supervision information, and is used to constrain the tensor decomposition process, thereby improving the accuracy and robustness of the traffic flow data completion method.

[0062] The terms "first", "second", "third", "fourth", etc. in the specification, claims and drawings of the present application are used to distinguish different objects, rather than to describe a specific order. In addition, the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but optionally further includes steps or units not listed, or optionally further includes other steps or units inherent to these processes, methods, products or devices.

[0063] Referring to "embodiments" herein means that specific features, structures, or characteristics described in connection with the embodiments can be included in at least one embodiment of the present application. The phrase appears in various places in the specification and does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment mutually exclusive with other embodiments. Those skilled in the art will explicitly and implicitly understand that the embodiments described herein can be combined with other embodiments.

[0064] "Plurality" means two or more. "And / or" describes the association relationship of associated objects and indicates that three relationships can exist. For example, A and / or B can represent: A exists alone, A and B exist simultaneously, and B exists alone. The character " / " generally represents an "or" relationship between the associated objects before and after.

[0065] The embodiments of the present application will be described below with reference to the accompanying drawings.

[0066] This embodiment provides a heterogeneity-guided traffic flow completion method, as Figure 1 shown, the method includes the following steps:

[0067] Step S101: Perform spatio-temporal decoupled mask encoding on the original traffic flow data to generate mask data , and the mask data includes temporal mask data and spatial mask data .

[0068] Specifically, in this embodiment, the traffic flow is defined as follows:

[0069] Traffic volume: Traffic volume refers to the number of vehicles passing through the road section where the monitoring point is located within a specific time interval. Then monitoring points within days and time intervals of traffic volume can be represented as a tensor .

[0070] Among them, the original traffic flow data is the traffic flow data with missing values, which is represented in tensor form to obtain the original traffic flow tensor .

[0071] Therefore, the problem definition of this embodiment is as follows: Given the traffic volume with missing values , the goal of this embodiment is to build a model to complete the missing values in.

[0072] Step by step, for the convenience of description, the symbol definitions of this embodiment are as follows:

[0073] Table 1 Symbol Definitions

[0074]

[0075] Step S102: Respectively perform feature embedding and positional encoding on the original traffic flow data and the mask data to obtain the spatio-temporal encoding results; among them, the spatio-temporal encoding results include the time and space encoding results of the original traffic flow data and the time and space encoding results of the mask data.

[0076] Through feature embedding, the original traffic flow data and the mask data are mapped to a high-dimensional feature space to extract their deep embedding features in the time dimension and the space dimension to capture the spatio-temporal characteristics of the data. By introducing a positional encoding mechanism, the time and space position encodings corresponding to the original traffic flow data and the mask data are constructed to capture their structured position information in the time dimension and the space dimension. By combining feature embedding and positional encoding to generate spatio-temporal encoding results, the model's ability to capture time periodic features and spatial context information can be enhanced, so as to achieve more efficient completion performance under different missing patterns.

[0077] Step S103: Based on the spatio-temporal encoding results, use the mask contrast learning method of reconstruction loss to generate a time heterogeneity representation matrix and a space heterogeneity representation matrix respectively.

[0078] Among them, the reconstruction loss is an objective function that measures the difference between the model's predicted value and the true value. In the data completion task, the reconstruction loss is used to evaluate the model's recovery effect on the missing part.

[0079] Mask learning enhances the model's understanding of the internal structure of data by masking (i.e., "masking") part of the information in the input data, enabling the model to learn how to infer the masked part from the remaining information. Contrastive learning is a self-supervised learning method whose core idea is to learn high-quality feature representations by maximizing the similarity of positive sample pairs and minimizing the similarity of negative sample pairs.

[0080] The masked contrastive learning method for reconstruction loss combines the ideas of mask learning and contrastive learning, which can generate high-quality feature representations, thereby improving the model's performance in filling in missing data without explicitly constructing positive and negative sample pairs. Instead, this method mainly relies on the mask reconstruction task itself to guide the model to learn useful feature representations.

[0081] Step S104: Using the temporal heterogeneity representation matrix and the spatial heterogeneity representation matrix as constraint conditions, complete the traffic flow data based on the tensor decomposition model to obtain the completed traffic flow .

[0082] Figure 2 shows the logic block diagram of the heterogeneity-guided traffic flow completion method proposed in this embodiment. As Figure 2 shown, the method includes a spatio-temporal decoupled mask encoding strategy, a spatio-temporal heterogeneity representation learning method for road traffic flow considering missing values, and a tensor decomposition method considering the spatio-temporal heterogeneity of road traffic flow. Among them, the spatio-temporal decoupled mask encoding strategy is used to perform mask setting and feature encoding separately from the time and space dimensions of the data to obtain the spatio-temporal encoding results of the data (including the original data, i.e., the time and space encodings of the original traffic flow data and the time and space encodings of the masked data); the spatio-temporal heterogeneity representation learning method for road traffic flow considering missing values is used to capture the spatio-temporal distribution pattern of the data from the embedded features of the traffic flow based on masked contrastive learning for reconstruction loss, thereby constructing a weight matrix representing the heterogeneity relationship between monitoring points and obtaining the heterogeneity representation results (including the temporal heterogeneity representation matrix and the spatial heterogeneity representation matrix); the tensor decomposition method considering the spatio-temporal heterogeneity of road traffic flow uses the spatio-temporal heterogeneity representation results to constrain the spatio-temporal relationship of the traffic flow, and optimizes the spatio-temporal representation process and the traffic flow completion process through self-supervised learning to enhance the completion performance of road traffic flow data.

[0083] The traffic flow completion task aims to complete the values at missing positions by establishing the spatio-temporal distribution pattern relationship between known data and unknown data. However, the diverse missing patterns of the data may interfere with the learning of the spatio-temporal distribution pattern, resulting in insufficient accuracy of the spatio-temporal heterogeneity representation, and further reducing the completion performance. Based on this, this embodiment proposes a spatio-temporal decoupled mask encoding strategy, including two parts: the generation of masked data and spatio-temporal encoding. By using the original traffic flow data Randomly set masks along the temporal and spatial dimensions to generate masked data and use a spatio-temporal encoder to generate and high-dimensional embedded features in the temporal and spatial dimensions, enabling the model to effectively evaluate the heterogeneity representation results of missing positions.

[0084] Among them, the generation of the masked data can be performed as follows:

[0085] Step S101a: According to the original traffic flow tensor , randomly generate a temporal mask tensor and a spatial mask tensor along the temporal dimension and the spatial dimension respectively at the mask rate .

[0086] Step S101b: Perform an element-wise multiplication operation on the original traffic flow tensor with the temporal mask tensor and the spatial mask tensor respectively to obtain the masked data .

[0087] Specifically, given the original traffic flow tensor , expand it along the spatial dimension and the temporal dimension respectively, and set the temporal and spatial masked data and as follows:

[0088] (1)

[0089] (2)

[0090] In the formula, and are matrices with only 0 and 1, randomly masks the time series of sensors, randomly masks the time series of

[0091] time steps; Figure 2 is the mask rate ranging from 0 to 1.

[0092] In some embodiments, feature embedding and position encoding are respectively performed on the original traffic flow data and the masked data to obtain the spatio-temporal encoding result, specifically:

[0093] Step S102a: Combine the block embedding technique and the fully connected layer to generate high-dimensional embedding features of the original traffic flow data and the mask data respectively;

[0094] Step S102b: Use the sine position encoding technique to generate the spatio-temporal position encodings of the original traffic flow data and the mask data respectively;

[0095] Step S102c: Element-wise multiply the high-dimensional embedding features of the time dimension and the space dimension of the original traffic flow data and the mask data with the spatio-temporal position encodings of the original traffic flow data and the mask data respectively, and then input them into the Transformer to obtain the spatio-temporal encoding result.

[0096] In this embodiment, the spatio-temporal encoding operation is performed separately by the time encoder and the space encoder. As Figure 2 shown, the time encoder and the space encoder are two parallel branches in the model, and they have similar structures. Both are composed of three parts: Patch Embedding, position encoding, and Transformer layer connected in sequence, and only the position encoding part is different: the time encoder is responsible for performing time position encoding, while the space encoder performs space position encoding. Through the time encoder and the space encoder, the self-attention mechanism is introduced along the time and space dimensions respectively, so that the model can effectively capture the long-range dependence relationship and local features in the data.

[0097] In this embodiment, in step S102a, the block embedding (Patch Embedding) technique is adopted. First, the original data and the mask data are cut into F patches according to the time length, and then each patch is converted into a representation form with a specific embedding dimension by a fully connected layer.

[0098] Specifically, the time dimension of the original data and the mask data is cut into according to the time length to obtain the patch inputs , and , ; Use the fully connected layer to obtain the embedding dimensions , and , of each patch input, where is the embedding dimension. The introduction of the block embedding technique is beneficial to enhancing the model's ability to capture complex and diverse features.

[0099] In steps S102b - S102c, sinusoidal positional encoding is introduced to learn and the time and spatial positional encoding (Time / Spatial Positional Encoding) of 、 and 、 . Subsequently, the positional encoding is combined with the feature embedding and input into the Transformer, i.e., the spatio - temporal positional encoding, to obtain the spatio - temporal encoding result 、 and 、 . The expression of the above process is as follows:

[0100] (3)

[0101] (4)

[0102] (5)

[0103] (6)

[0104] (7)

[0105] In the formula, 、 respectively represent the high - dimensional embedding features in the time dimension and the spatial dimension, 、 respectively represent the high - dimensional embedding features in the time dimension and the spatial dimension; 、 respectively represent the time encoding and the spatial positional encoding; 、 respectively represent the time encoding and the spatial positional encoding, is the embedding dimension, represents the element - wise multiplication operation, represents the index of the spatial position, represents the index of the feature dimension, and are used to control the frequency of the piece - wise sine / cosine function, enhance the diversity of the positional encoding, and avoid frequency repetition between different dimensions.

[0106] By combining the positional encoding with the feature embedding, the model can more effectively utilize the time - periodic features and spatial context information, thereby improving its completion performance under different missing patterns.

[0107] Considering that the key to spatio-temporal heterogeneity representation learning lies in accurately mining the spatio-temporal distribution patterns of traffic flow to identify and understand its dynamic changes and spatial differences. Existing spatio-temporal heterogeneity representation learning methods usually assume that the data is complete, while in practical applications, data missing is inevitable, which may lead to incomplete acquisition of traffic flow patterns by the model, thus affecting the accuracy and reliability of spatio-temporal heterogeneity representation. Therefore, this embodiment adopts a spatio-temporal heterogeneity representation learning method for road traffic flow considering missing values to obtain spatio-temporal heterogeneity representation results (i.e., temporal heterogeneity representation matrix and spatial heterogeneity representation matrix), so as to effectively learn and capture clearer heterogeneity representation in the case of data missing.

[0108] Specifically, based on the spatio-temporal encoding results, using the masked contrastive learning method of reconstruction loss, generate the temporal heterogeneity representation matrix and the spatial heterogeneity representation matrix respectively, including:

[0109] Step S103a: Perform mean and dimension conversion operations on the temporal encoding results of the original traffic flow data and the masked data respectively to extract the average feature representation of each time step of the original traffic flow data and the masked data 、 。

[0110] Step S103b: Introduce the first parameter matrix 、the second parameter matrix , and perform multiplication operations on each time dimension of the first parameter matrix 、the second parameter matrix and the average feature representations 、 of the original traffic flow data and the masked data respectively to obtain the temporal heterogeneity representation results 、 of the original traffic flow data and the masked data respectively.

[0111] Step S103c: Calculate the contrastive loss between and to obtain the first loss 。

[0112] Step S103d: Perform mean operations on the spatial encoding results of the original traffic flow data and the masked data respectively in the embedding dimension to extract the average feature representations 、 of the original traffic flow data and the masked data at each node position respectively.

[0113] Step S103e: Introduce the third parameter matrix 、the fourth parameter matrix , and... , the fourth parameter matrix and the average feature representations of the original traffic flow data and the mask data at each node position respectively , perform a multiplication operation to obtain a spatial heterogeneity characterization result , .

[0114] Step S103f: Calculate and 's contrast loss to obtain the second loss .

[0115] Step S103g: Through iterative optimization of the first loss , the second loss , to generate the final temporal heterogeneity characterization matrix , and the spatial heterogeneity characterization matrix , .

[0116] Exemplarily, the above steps S103a to S103g can be specifically executed as follows:

[0117] For the time encoding results and , perform mean and dimension transformation operations on their embedding dimensions respectively to extract the average feature representations of each time step and , and the formula is as follows:

[0118] (8)

[0119] (9)

[0120] In the formula, , , , represents the result after the mean and dimension transformation operations for each time step.

[0121] To effectively capture the feature differences of different monitoring points in the time dimension and enhance the model's learning ability for temporal heterogeneity features, this embodiment introduces learnable first parameter matrix , the second parameter matrix , where . Multiply the first parameter matrix , the second parameter matrix respectively with and for each time dimension to obtain the temporal heterogeneity characterization results and , the expression is as follows:

[0122] (10)

[0123] (11)

[0124] In the formula, represents the slicing extraction operation, represents the transpose operation.

[0125] On this basis, a method of masked contrastive learning based on reconstruction loss is used to calculate the and contrast loss between , that is, the first loss is obtained, and the expression is as follows:

[0126] (12)

[0127] Similarly, for the spatial encoding results of the original traffic flow data and the masked data , , the spatial heterogeneity characterization process is respectively performed. That is, the mean operation is performed on their embedding dimensions to extract the average feature representation of each node position , , and the expression is as follows:

[0128] (13)

[0129] (14)

[0130] Similarly, in order to effectively capture the feature differences of different monitoring points in the spatial dimension and enhance the model's learning ability of spatial heterogeneity features, this embodiment introduces a third parameter matrix , a fourth parameter matrix , where , and , are respectively multiplied by , to obtain the spatial heterogeneity characterization results and , and the expression is as follows:

[0131] (15)

[0132] (16)

[0133] Using the method of masked contrastive learning of reconstruction loss, calculate and Contrastive loss between them That is, the second loss is obtained :

[0134] (17)

[0135] Finally, by continuously iterating and optimizing the above loss function, the model can effectively update the learning parameters, thereby enhancing the model's ability to capture spatial heterogeneity features. Based on the iterated learning parameters, the final temporal heterogeneity representation matrix 、 and the spatial heterogeneity representation matrix 、 .

[0136] The tensor decomposition model effectively simplifies the complex high-dimensional data structure into an easy-to-process low-dimensional representation by decomposing the high-dimensional tensor into multiple low-dimensional factors, thereby extracting the latent features and hidden patterns in the data. This characteristic makes the tensor decomposition method show strong applicability in dealing with the problem of urban traffic flow completion. Urban traffic flow data usually has significant high-dimensional characteristics, involving multiple dimensions such as time and space. The tensor decomposition method can represent these high-dimensional data as a set of low-dimensional factors to identify and extract the latent features in the traffic flow. However, the existing methods mainly complete the missing data based on the spatio-temporal correlation of the traffic flow data, often ignoring its inherent spatio-temporal heterogeneity features, thus affecting the accuracy and robustness of the completion result. In view of this, this embodiment adopts a tensor decomposition method that takes into account spatio-temporal heterogeneity, and enhances the constraint on the spatio-temporal relationship of the traffic flow by embedding spatio-temporal heterogeneity representations, thereby improving the performance and robustness of data completion.

[0137] Specifically, in this embodiment, the tensor decomposition model is a CP decomposition model. The original traffic flow tensor is decomposed into the first factor matrix , the second factor matrix , and the third factor matrix , that is, the original traffic flow tensor is expressed as a multilinear combination of latent factors in each factor matrix. The expression is as follows:

[0138] (18)

[0139] In the formula, is an element of the tensor , , and are the indices of the three dimensions, is a global parameter used to approximate the overall average value of the tensor elements, , and are factor matrices that control the interactions between different dimensions. is the serial number of the latent factor.

[0140] To address the non-convex optimization problem of tensor decomposition, this embodiment introduces the probability distribution of the factor matrix through variational Bayesian learning, that is, optimizes its probability distribution through variational Bayesian learning. Specifically, first let the variational posterior parameter follow a Gaussian distribution .

[0141] Secondly, in order to be able to use the gradient descent method for optimization, the reparameterization trick is used to initialize the weights of the first factor matrix , the second factor matrix , and the third factor matrix : , and :

[0142] (19)

[0143] In the formula, , , are the weights of the first factor matrix , the second factor matrix , and the third factor matrix respectively. , , are the global parameters corresponding to the first factor matrix , the second factor matrix , and the third factor matrix respectively. is the standard deviation of the variational posterior parameter . follows a Gaussian distribution. , , are the random noises corresponding to the first factor matrix , the second factor matrix , and the third factor matrix respectively. , , all follow a normal distribution, that is , and are random variables.

[0144] Among them, to ensure it is always non-negative.

[0145] In step S103, the time heterogeneity characterization matrix and the spatial heterogeneity characterization matrix are used as the constraint conditions of the tensor decomposition model, that is, the and of the spatial heterogeneity characterization results are used to constrain the process of tensor decomposition. Then, the loss function can be expressed as the sum of four parts: the reconstruction error term, the spatial heterogeneity characterization constraint term, the time heterogeneity characterization constraint term, and the regularization term. Among them,

[0146] The reconstruction error term is the reconstruction error between the original traffic flow and the completed traffic flow.

[0147] The spatial heterogeneity characterization constraint term is used to constrain the spatial relationship of traffic flows between different nodes, and its expression is:

[0148] (20)

[0149] In the formula, is the spatial heterogeneity characterization matrix of the original traffic flow data; is the first factor matrix;

[0150] The time heterogeneity characterization constraint term is used to constrain the time relationship of traffic flows between different nodes, and its expression is as follows:

[0151] (21)

[0152] In the formula, is the time heterogeneity characterization matrix of the original traffic flow data, , are the elements of the first factor matrix with indices ( ), ( ); is the element of the second factor matrix with index ( );

[0153] The expression of the regularization term is:

[0154] (22)

[0155] Summing up the above reconstruction error term, spatial heterogeneity characterization constraint term, time heterogeneity characterization constraint term, and regularization term, the expression of the loss function is as follows:

[0156] (23)

[0157] In the formula, is the reconstruction error term, which is used to control the decomposition error, is the spatial heterogeneity characterization constraint term, which controls the factor The decomposition error is used to constrain the spatial relationship of traffic flow between different nodes. is the constraint term for time heterogeneity characterization, and the control factor The decomposition error is used to constrain the temporal relationship of traffic flow between different nodes. is the regularization term to prevent overfitting of the objective function; , and are the weights for controlling the regularization term; is the completion result for each round.

[0158] Furthermore, the above process is implemented in pseudocode as follows:

[0159]

[0160] In the above code implementation, first initialize the three factor matrices for traffic flow decomposition , and , and use the masked data of and in the time and space dimensions as the inputs of spatio-temporal encoding respectively to obtain the spatio-temporal encoded data , , and (see lines 1 - 6 of the code); then, execute the time heterogeneity characterization module for and respectively, and obtain the loss values of and based on masked contrastive learning of the reconstruction loss , and update the parameters of the time heterogeneity characterization module (see lines 7 - 13 of the code); then, execute the spatial heterogeneity characterization module for and respectively, and also obtain the loss values of and based on masked contrastive learning of the reconstruction loss , and update the parameters of the spatial heterogeneity characterization module (see lines 14 - 19 of the code); based on this, embed and into the tensor decomposition model as constraint conditions to control the decomposition error of the objective function (see lines 20 - 24 of the code). The above process is iterated continuously, calculating the error between the completion value and the original tensor, and backpropagating the loss value to each module of spatio-temporal heterogeneity characterization and factor matrix solution, so as to optimize the spatio-temporal heterogeneity characterization and tensor decomposition process (see lines 25 - 26 of the code).

[0161] ​​In addition, to verify the performance of the model, experiments were conducted on a real collected urban road traffic dataset and compared with the state-of-the-art baseline model. The results show that the method proposed in this embodiment can effectively characterize the spatio-temporal heterogeneity of road traffic flow data under different missing patterns and exhibit high flow data completion accuracy and robustness.

[0162] In summary, the method proposed in this application combines self-supervised learning technology with tensor decomposition, effectively represents spatio-temporal heterogeneity through a heterogeneity representation matrix and a spatial heterogeneity representation matrix, and uses spatio-temporal heterogeneity-guided tensor decomposition to improve the accuracy of traffic flow data completion methods.

[0163] By calculating the spatio-temporal heterogeneity representation through a spatio-temporal heterogeneity representation learning method for road traffic flow considering missing values, the ability of the model to reconstruct the spatio-temporal heterogeneity at the missing positions can be enhanced; using the spatio-temporal heterogeneity representation results to constrain the spatio-temporal relationship of traffic flow enhances the completion performance of road traffic flow data.

[0164] Based on the same inventive concept, this embodiment provides a heterogeneity-guided traffic flow completion system, which includes a mask encoding module, an embedding and position encoding module, a spatio-temporal heterogeneity representation module, and a decomposition module. Among them:

[0165] The mask encoding module is used to perform spatio-temporal decoupled mask encoding on the original traffic flow data to generate mask data , and the mask data includes time mask data and spatial mask data ; where the original traffic flow data is traffic flow data with missing values, which is represented in tensor form to obtain the original traffic flow tensor ;

[0166] The embedding and position encoding module is used to perform feature embedding and position encoding on the original traffic flow data and the mask data respectively to obtain spatio-temporal encoding results; where the spatio-temporal encoding results include the time and space encoding results of the original traffic flow data and the mask data respectively;

[0167] The spatio-temporal heterogeneity representation module is used to generate a time heterogeneity representation matrix and a spatial heterogeneity representation matrix respectively based on the spatio-temporal encoding results by using a mask contrast learning method of reconstruction loss;

[0168] The decomposition module is used to use the time heterogeneity representation matrix and the spatial heterogeneity representation matrix as constraint conditions to complete the traffic flow data based on the tensor decomposition model to obtain the completed traffic flow .

[0169] The heterogeneous-guided traffic flow completion system provided in this embodiment can implement the steps and processes of the heterogeneous-guided traffic flow completion method provided in any of the above embodiments and achieve the same technical effects, which will not be elaborated here one by one.

[0170] The embodiments of this application can be applied to Figure 3 the electronic devices shown in the figure. The electronic devices can be, but are not limited to, mobile terminals such as mobile phones, tablet computers, handheld computers, personal digital assistants (PDAs), etc., smart home devices such as smart TVs and smart cameras, wearable devices such as smart bracelets, smart watches, and smart glasses, or other computer devices such as desktop computers, laptop computers, ultra-mobile personal computers (UMPCs), netbooks, and smart screens.

[0171] As Figure 3 shown in the figure, the electronic device 200 may include one or more of the following components: a processor 201, a memory 203, a communication interface 202, and a communication bus 204. Among them, the memory 203 can be connected to the processor 201 through the bus 204. The bus can transmit data between the processor 201 and the memory 203. The bus can be divided into an address bus, a data bus, a control bus, etc.

[0172] The processor 201 may include one or more processing cores. The processor 201 can connect various parts within the entire electronic device 200 using various interfaces and circuits. By running or executing instructions, programs, code sets, or instruction sets stored in the memory 203, and by invoking the data stored in the memory 203, it performs various functions of the electronic device 200 and processes data. Exemplarily, the processor 201 may include an application processor (AP), a modem processor, a CPU, a graphics processing unit (GPU), an image signal processor (ISP), a controller, a video codec, a digital signal processor (DSP), a field-programmable gate array (FPGA), a programmable logic array (PLA), and / or a neural-network processing unit (NPU), etc. Among them, the CPU mainly processes the operating system, user interface, and application programs, etc.; the GPU is responsible for rendering and drawing the content to be displayed; the NPU is used to implement artificial intelligence (AI) functions; the modem is used to handle wireless communications. Different processing units can be independent devices or integrated in one or more processors. For example, the multiple processing units shown above are all integrated in one SoC, or the AP is a separate semiconductor chip and other processing units are integrated in one SoC. This application does not make any limitations in this regard.

[0173] The memory 203 may include a random access memory (RAM), may also include a read-only memory (ROM), and may further include a non-transitory computer-readable storage medium. The memory 203 can be used to store instructions, programs, code, code sets, or instruction sets. The memory 203 may include a program storage area and a data storage area. Among them, the program storage area can store instructions for implementing the operating system, instructions for at least one function, such as the heterogeneous-guided traffic flow completion method, etc.; the data storage area can store data created according to the use of the electronic device 200, such as road traffic flow data, intermediate model output results, etc.

[0174] In addition, those skilled in the art can understand that the structure of the electronic device 200 shown in the above drawings does not limit the electronic device 200. The electronic device may include more or fewer components than those shown, or combine some components, or have different component arrangements. For example, the electronic device 200 further includes components such as a microphone, a speaker, a radio frequency circuit, a sensor, an audio circuit, a power supply, and a Bluetooth module, which will not be elaborated here.

[0175] The embodiment of the present application also provides a computer program product including computer-executable instructions. In one embodiment, the computer-executable instructions are used to cause a computer to execute the functions in the above method embodiment.

[0176] The computer-executable instructions can be stored in a computer-readable storage medium. The embodiment of the present application also provides a computer-readable storage medium, in which executable instructions are stored. In one embodiment, the computer-executable instructions are used to cause a computer to execute the functions in the above method embodiment.

[0177] The above are only the preferred embodiments of the present application and are not used to limit the present application. For those skilled in the art, the present application can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A heterogeneity-guided traffic flow completion method, characterized in that: The following steps are involved: Perform spatiotemporal decoupling mask encoding on the original traffic flow data to generate mask data , the mask data includes time mask data and spatial mask data ; The original traffic flow data is traffic flow data with missing values, which is expressed as a tensor to obtain the original traffic flow tensor ; Performing feature embedding and position encoding on the original traffic flow data and the mask data respectively to obtain a spatiotemporal encoding result; wherein the spatiotemporal encoding result includes the time and space encoding results of the original traffic flow data and the time and space encoding results of the mask data; Based on the spatiotemporal encoding result, a mask contrast learning method with reconstruction loss is used to generate a temporal heterogeneity representation matrix and a spatial heterogeneity representation matrix respectively; The time heterogeneity characterization matrix and the space heterogeneity characterization matrix are used as constraints, and the traffic flow data is completed based on the tensor decomposition model to obtain the completed traffic flow. ; Based on the spatiotemporal coding result, a mask contrast learning method with reconstruction loss is used to generate a temporal heterogeneity representation matrix and a spatial heterogeneity representation matrix, respectively, including: The time encoding results of the original traffic flow data and the mask data are respectively subjected to mean and dimension conversion operations to extract the average feature representation of each time step of the original traffic flow data and the mask data. , ; Introducing the first parameter matrix , the second parameter matrix , and the first parameter matrix , the second parameter matrix The average feature values ​​of the original traffic flow data and the mask data are shown in Figure 2. , A multiplication operation is performed on each time dimension to obtain the temporal heterogeneity characterization results of the original traffic flow data and the mask data. , ; calculate and The contrast loss is obtained by ; In the embedding dimension, the spatial encoding results of the original traffic flow data and the mask data are respectively averaged to extract the average feature representation of the original traffic flow data and the mask data at each node position. , ; Introducing the third parameter matrix , the fourth parameter matrix , and the third parameter matrix , the fourth parameter matrix The average feature representation of the original traffic flow data and the mask data at each node position is , Perform multiplication to obtain spatial heterogeneity characterization results , ; calculate and The contrast loss is obtained, and the second loss is obtained ; By iteratively optimizing the first loss The second loss , to generate the final temporal heterogeneity representation matrix , and spatial heterogeneity characterization matrix , .

2. The method according to claim 1, characterized in that: Perform spatiotemporal decoupling mask encoding on the original traffic flow data to generate mask data ,include: According to the original traffic flow tensor , along the time dimension and the space dimension respectively according to the mask rate Randomly generate time mask tensors and the spatial mask tensor ; The original traffic flow tensor With the time mask tensor , spatial mask tensor Perform element multiplication operations respectively to obtain the mask data .

3. The method according to claim 1, characterized in that: The original traffic flow data and the mask data are respectively subjected to feature embedding and position encoding to obtain a spatiotemporal encoding result, which is specifically: Using a combination of block embedding technology and a fully connected layer to generate high-dimensional embedding features of the original traffic flow data and the mask data respectively; The sinusoidal position coding technique is used to generate the spatiotemporal position coding of the original traffic flow data and the mask data respectively; The high-dimensional embedding features of the original traffic flow data and the mask data are element-wise multiplied with the spatiotemporal position codes of the original traffic flow data and the mask data, and then input into the Transformer to obtain the spatiotemporal coding result.

4. The method according to claim 1, characterized in that: The tensor decomposition model is a CP decomposition model, which uses the CP decomposition model to decompose the original traffic flow tensor into a first factor matrix , the second factor matrix , the third factor matrix , and optimize its probability distribution through variational Bayesian learning.

5. The method according to claim 4, characterized in that The loss function of the CP decomposition model is the sum of the following four parts: reconstruction error term, spatial heterogeneity representation constraint term, temporal heterogeneity representation constraint term and regularization term; The reconstruction error term is the reconstruction error between the original traffic flow and the completed traffic flow; The spatial heterogeneity characterization constraint term is used to constrain the spatial relationship of traffic flow between different nodes, and its expression is: ， In the formula, is a spatial heterogeneity representation matrix of the original traffic flow data; is the first factor matrix; The temporal heterogeneity characterization constraint term is used to constrain the temporal relationship of traffic flow between different nodes, and its expression is as follows: ， In the formula, is the temporal heterogeneity characterization matrix of the original traffic flow data, , is the first factor matrix The index is ( )、( ), is the second factor matrix The index is ( ) elements; The expression of the regularization term is: .

6. The method according to claim 4, characterized in that The first factor matrix , the second factor matrix , the third factor matrix The weights of are initialized using the reparameterization technique, specifically: ， ， ， In the formula, , , The first factor matrix is , the second factor matrix , the third factor matrix The weight of , , The first factor matrix is , the second factor matrix , the third factor matrix The corresponding global parameters, is the variational posterior parameter The standard deviation of It conforms to the Gaussian distribution. , , The first factor matrix is , the second factor matrix , the third factor matrix The corresponding random noise is , , All obey normal distribution.

7. A heterogeneity-guided traffic flow completion system, characterized in that: include: The mask encoding module is used to perform spatiotemporal decoupling mask encoding on the original traffic flow data to generate mask data. , the mask data includes time mask data and spatial mask data ; The original traffic flow data is traffic flow data with missing values, which is expressed as a tensor to obtain the original traffic flow tensor ; An embedding and position coding module, used to perform feature embedding and position coding on the original traffic flow data and the mask data respectively, to obtain a spatiotemporal coding result; wherein the spatiotemporal coding result includes the time and space coding results of the original traffic flow data and the mask data respectively; A spatiotemporal heterogeneity representation module, for generating a temporal heterogeneity representation matrix and a spatial heterogeneity representation matrix respectively based on the spatiotemporal encoding result and using a mask contrast learning method with reconstruction loss; A decomposition module is used to use the temporal heterogeneity representation matrix and the spatial heterogeneity representation matrix as constraints, and to complete the traffic flow data based on the tensor decomposition model to obtain the completed traffic flow. ; The spatiotemporal heterogeneity characterization module is further configured as follows: The time encoding results of the original traffic flow data and the mask data are respectively subjected to mean and dimension conversion operations to extract the average feature representation of each time step of the original traffic flow data and the mask data. , ; Introducing the first parameter matrix , the second parameter matrix , and the first parameter matrix , the second parameter matrix The average feature values ​​of the original traffic flow data and the mask data are shown in Figure 2. , A multiplication operation is performed on each time dimension to obtain the temporal heterogeneity characterization results of the original traffic flow data and the mask data. , ; calculate and The contrast loss is obtained by ; In the embedding dimension, the spatial encoding results of the original traffic flow data and the mask data are respectively averaged to extract the average feature representation of the original traffic flow data and the mask data at each node position. , ; Introducing the third parameter matrix , the fourth parameter matrix , and the third parameter matrix , the fourth parameter matrix The average feature representation of the original traffic flow data and the mask data at each node position is , Perform multiplication to obtain spatial heterogeneity characterization results , ; calculate and The contrast loss is obtained, and the second loss is obtained ; By iteratively optimizing the first loss The second loss , to generate the final temporal heterogeneity representation matrix , and spatial heterogeneity characterization matrix , .

8. An electronic device, characterized in that: include: a memory for storing instructions executed by one or more processors of the electronic device; The processor, when the processor executes the instructions in the memory, can enable the electronic device to implement the steps of the heterogeneity-guided traffic flow completion method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores instructions, which, when executed on a computer, implement the steps of the heterogeneity-guided traffic flow completion method as described in any one of claims 1 to 6.

Images