Traffic missing data complementation method based on C-Tucker fourth-order space-time tensor representation

Abstract

The invention discloses a traffic missing data complementing method based on C-Tucker fourth-order space-time tensor representation, which comprises the following steps: constructing a fourth-order space-time tensor model of four dimensions of time, day, period and space based on space-time relevance of traffic flow data; performing Tucker decomposition on the observation tensor to calculate a core tensor; n-order modal expansion of the observation tensor and the core tensor is calculated, singular value decomposition (SVD) is adopted to extract a feature vector set, and a factor moment Nt and a core matrix are formed based on feature vectors to construct a fusion matrix; and reconstructing the observation tensor through the core tensor and the fusion matrix to realize recovery of traffic data. According to the method, under the condition that discrete data and continuous missing data coexist, efficient and robust data completion is achieved, and the accuracy of data recovery is effectively improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of traffic data completion, and particularly relates to a method for completing missing traffic data based on a C-Tucker fourth-order spatio-temporal tensor representation. Background Art

[0002] In the field of intelligent transportation, the integrity of traffic data directly affects the accuracy of road network planning and congestion prediction. However, due to factors such as sensor failures and communication errors, the multi-source traffic data actually collected often has discrete random missing (such as single-point instantaneous failures) and continuous random missing (such as multi-week data interruptions), and the two types of missing often coexist. The current traffic data completion technologies mainly adopt algorithms such as historical data interpolation, neighboring data averaging, and matrix decomposition-based (such as PCA, BPCA, PPCA).

[0003] Although these technologies have solved the problem of completing data of a single missing type to a certain extent, they still have the following deficiencies: First, the existing methods mainly focus on the correlation between time series or local data, and fail to fully exploit the inherent global spatial and temporal periodic characteristics in traffic data, resulting in difficulty in capturing and utilizing the potential spatio-temporal correlation of traffic flow in the case of data missing at multiple detection points and continuous time periods. Second, most traditional technologies are mainly designed for discrete missing data. When discrete and continuous missing data coexist, their completion effect is significantly reduced, and the completion requirements of the two missing modes cannot be satisfied simultaneously. Summary of the Invention

[0004] Object of the Invention: The object of the present invention is to provide a method for completing missing traffic data based on a C-Tucker fourth-order spatio-temporal tensor representation. By constructing a fourth-order spatio-temporal tensor model with "time, day, period, and location" as modes, and combining singular value decomposition to optimize the factor matrices, efficient and robust data completion is achieved in the case of coexistence of discrete and continuous missing data, effectively improving the accuracy of data recovery.

[0005] Technical Solution: A method for completing missing traffic data based on a C-Tucker fourth-order spatio-temporal tensor representation of the present invention includes the following steps:

[0006] Step 1: Based on the spatio-temporal correlation of traffic flow data, construct a fourth-order spatio-temporal tensor model including four dimensions of time, day, period, and space, and initialize the observed tensor X;

[0007] Step 2: Based on the fourth-order spatio-temporal tensor model, perform Tucker decomposition on the observed tensor X to calculate the core tensor G;

[0008] Step 3: Calculate the n - order mode expansion of the observed tensor X and the core tensor G, extract the set of eigenvectors using singular value decomposition (SVD), and select the eigenvectors with the largest eigenvalues according to the dimension of each order to form the factor matrix N t and the core matrix M t ;

[0009] Step 4: Based on the eigenvectors, form the factor matrix N t and the core matrix M t to construct the fusion matrix R t ;

[0010] Step 5: Reconstruct the observed tensor X through the core tensor G and the fusion matrix R t to achieve the recovery of traffic data.

[0011] Furthermore, Step 1 is specifically as follows: Based on the spatio - temporal correlation of traffic flow data, construct a fourth - order spatio - temporal tensor model including the following dimensions:

[0012] Time, I1 = 288: With a 5 - minute time interval, each day is divided into 288 time slices to capture the high and low fluctuations of intraday traffic;

[0013] Day, I2 = 28: Taking 4 weeks (i.e., 28 days) as a single cycle to model the traffic flow changes and their correlations between different days;

[0014] Cycle, I3 = 13: The 52 - week year is divided into 13 cycles, each cycle containing 4 - week data, used to explicitly encode the long - term patterns across months and seasons;

[0015] Location, I4 = 13: Using 13 spatially adjacent detection points to improve the robustness during completion through proximity constraints;

[0016] The complete traffic flow data is represented as a fourth - order tensor A ∈ R 288×28×13×13 , where each dimension represents time, day, cycle, and location in sequence, and it contains 1,362,816 elements in total.

[0017] Furthermore, Step 2 is specifically as follows: Construct a binary mask tensor W i,j,k,l ∈ {0, 1} 288×28×13×13 , and the expression is as follows:

[0018]

[0019] Then, obtain the observed tensor X through the Hadamard product operation:

[0020]

[0021] where, Denotes the Hadamard product operation symbol. A is the complete data set. In the operation result, the missing data is handled as 0, and the known data is retained, realizing the separation of the missing and observed parts in traffic data;

[0022] Perform Tucker decomposition on the observed tensor X to obtain the core tensor G and the factor matrices U of each order t , t ∈ [1:4]:

[0023] X ≈ G ×1U1 ×2U2 ×3U3 ×4U4

[0024] Among them, the core tensor Stores the global correlation features across dimensions and satisfies K t <I t , t ∈ [1:4]; The factor matrix Represents the local patterns of each dimension respectively.

[0025] Furthermore, step 3 is specifically as follows: First, expand the observed tensor X and the core tensor G along each order of modality respectively, denoted as:

[0026] X t =Γ(X, t), G t =Γ(G, t), t ∈ [1:4]

[0027] Among them, the functions Γ(X, t) and Γ(G, t) represent the t-order modality expansion of X and G; The matrix Is used to convert the high-order tensor into a two-dimensional matrix for matrix decomposition, Is used to extract the global correlation features of the core tensor at each order; Taking the first-order expansion matrix X1 ∈ R 288×(28×13×13) As an example, it represents the time dimension expansion matrix;

[0028] Perform SVD decomposition on the expansion matrices X t And G t :

[0029]

[0030] Among them, Is the left singular vector matrix of the matrix X t , with the dimension of I t ×I t ; Its column vectors correspond to the left singular vectors of X t , used to capture the local spatio-temporal patterns of the original data at the t-order modality; Is the diagonal singular value matrix, with the dimension of I t ×Π k≠t I k ; The diagonal elements are arranged in descending order, reflecting the importance of each singular vector; Is the matrix Xt The right singular vector matrix, with dimension Π k≠t I k ×Π k≠t I k ; its column vectors correspond to the right singular vectors of X t and assist in completing the orthogonal decomposition of the matrix; is the left singular vector matrix of matrix G t with dimension K t ×K t ; its column vectors are the global correlation patterns of the core tensor at the t-th order, used to guide the global fusion of local features; is the diagonal singular value matrix, with dimension K t ×Π k≠t K k ; the diagonal elements are arranged in descending order, reflecting the global feature importance of the core tensor at each order; is the right singular vector matrix of matrix G t with dimension Π k≠t K k ×Π k≠t K k ; its column vectors are the right singular vectors of the core tensor at the t-th order, participating in the orthogonal decomposition of global features;

[0031] In singular value decomposition, the magnitude of each singular value reflects the information energy carried by its corresponding singular vector. Select the first k t vectors corresponding to the largest singular values to construct the local feature matrix Meanwhile, select the first k t vectors to construct the global feature matrix

[0032] Furthermore, step 4 is specifically: construct the fused feature factor matrix through matrix multiplication:

[0033]

[0034] where N t represents the local feature matrix, and M t represents the global feature matrix.

[0035] Furthermore, step 5 is specifically: use the core tensor G and the fused factor matrices R1, R2, R3, R4 to reconstruct the complete tensor, and the calculation formula is:

[0036]

[0037] It is a tensor reconstructed by C-Tucker, which integrates the local and global features of each order of modality to simultaneously repair discrete missing and continuous missing data;

[0038] Introduce nuclear norm constraint to further ensure the low-rank structure, and obtain the objective function:

[0039]

[0040] where, is the Frobenius norm, W is the missing mask tensor, the parameter λ is used to control the weight of the low-rank constraint, and ||X t || represents the nuclear norm of the matrix after the observation tensor X is unfolded along the t-th order modality.

[0041] The present invention also discloses a computer device, including a memory, a processor, and a computer program stored on the memory, and the processor executes the computer program to implement the steps of the method of the present invention.

[0042] The present invention also discloses a computer-readable storage medium, on which a computer program / instructions are stored, and when the computer program / instructions are executed by a processor, the steps of the method of the present invention are implemented.

[0043] The present invention also discloses a computer program product, including a computer program / instructions, and when the computer program / instructions are executed by a processor, the steps of the method of the present invention are implemented.

[0044] Beneficial effects: Compared with the prior art, the present invention has the following remarkable advantages:

[0045] 1. The present invention innovatively constructs traffic flow data into a fourth-order out-of-control tensor of "time-day-cycle-location", which explicitly encodes the intra-day fluctuations, intra-week differences, cross-month patterns, and spatial proximities, overcomes the problem of insufficient expression of long-period correlations by traditional third-order models, and provides richer redundant information for data recovery in high missing rate scenarios;

[0046] 2. After completing the Tucker decomposition, the present invention unfolds the tensor X and the core tensor G respectively in terms of modality, and uses singular value decomposition to select the local and global eigenvectors corresponding to the first k t largest singular values, and constructs a fusion factor matrix R t through matrix multiplication, significantly improving the completion accuracy.

[0047] 3. The present invention aims at a complex scenario where discrete random missing (single-point failure) and continuous random missing (multi-week interruption) coexist, repairs these two types of missing at the same time, and effectively suppresses the interference of discrete missing on continuous missing, maintaining excellent data supplementation effect. Description of the Drawings

[0048] Figure 1 Schematic diagram of the structure of 13 detection points of the fourth-order spatio-temporal tensor of the present invention;

[0049] Figure 2 Schematic diagram of the structure of a single detection point of the fourth-order spatio-temporal tensor of the present invention;

[0050] Figure 3 Flowchart of a traffic missing data completion method based on the C-Tucker fourth-order spatio-temporal tensor representation of the present invention;

[0051] Figure 4 Effect comparison of the RMSE evaluation index between the present invention and the comparative method when missing data appears at different detection points;

[0052] Figure 5 Effect comparison of the R-Square evaluation index between the present invention and the comparative method when missing data appears at different detection points;

[0053] Figure 6 Completion performance comparison between the present invention and the comparative method under different missing rate conditions. Detailed implementation manners

[0054] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0055] Embodiment

[0056] Please refer to Figure 1 As shown in the schematic diagram of the structure of 13 detection points of the fourth-order spatio-temporal tensor, for the situation where discrete and continuous random missing data coexist, the present invention constructs a fourth-order spatio-temporal tensor model, which fully utilizes the inherent characteristics of traffic data in the time, day, cycle, and space dimensions, and simultaneously completes these two types of missing data. First, by performing Tucker decomposition on the observed tensor X, its core tensor G is obtained. Subsequently, X and G are respectively expanded according to each order, and the singular value decomposition (SVD) is used to extract the set of eigenvectors. According to the dimension of each order, the eigenvectors with the largest eigenvalues are selected to form the factor matrix N t and the core matrix M t , and these two matrices can better reflect the potential characteristics of the data than the matrices directly obtained by Tucker decomposition. To further capture the potential correlations between each order, a fusion matrix R t is constructed, which is obtained by the product of N t and M t . Finally, the core tensor G and the eigenmatrix R t are used to reconstruct the tensor data, thereby realizing the recovery of the complete data set. The present invention provides a technical solution:

[0057] Based on the inherent spatio-temporal correlation in traffic data, a fourth-order spatio-temporal tensor model is constructed It includes the following dimensions:

[0058] Time (I1 = 288): With a 5 - minute time interval, a day is divided into 288 time slices, which are used to capture the high and low fluctuations of intraday traffic.

[0059] Day (I2 = 28): Taking 4 weeks (28 days) as a single cycle, it is used to model the traffic flow changes and their correlations between different days.

[0060] Cycle (I3 = 13): The 52 - week whole year is divided into 13 cycles, and each cycle contains 4 - week data, which is used to explicitly encode the long - term regularities across months and quarters.

[0061] Location (I4 = 13): Using 13 spatially adjacent detection points, the robustness during completion is improved through proximity constraints.

[0062] Thus, the complete traffic flow data can be represented as a fourth - order tensor A ∈ R 288×28×13×13 , where each dimension represents time, day, cycle, and location in sequence, and it contains 1,362,816 elements in total. The present invention can not only capture intraday details, but also model long - term correlations through the cycle dimension, fully express the regularities across months and quarters, and provide redundant information to assist data recovery in high - missing scenarios.

[0063] The present invention realizes the collaborative completion of discrete and continuous missing values in traffic data through tensor decomposition and feature fusion. To ensure that only known data is used in the data completion process, a binary mask tensor W i,j,k,l ∈ {0, 1} 288×28×13×13 is constructed, and the expression is as follows:

[0064]

[0065] Then, the observed tensor X is obtained through the Hadamard product operation:

[0066]

[0067] where A is the complete data set, and the missing data in the operation result is disposed as 0, while the known data is retained, thus realizing the effective separation of the missing and observed parts in traffic data.

[0068] The Tucker decomposition is performed on the observed tensor X to obtain the core tensor G and factor matrices Ut t , t ∈ [1:4]:

[0069] X ≈ G × 1U1 × 2U2 × 3U3 × 4U4

[0070] where the core tensor stores the global correlation features across dimensions and satisfies Kt <I t , where \(t\in[1:4]\). The factor matrices represent the local patterns in each dimension respectively. For example, the column vectors of \(U_1\) encode the intraday volatility basis functions in the time dimension (such as the morning peak pattern), while the row vectors of \(U_3\) are used to model the cross-month correlations in the periodic dimension.

[0071] The accuracy of the factor matrices directly affects the data completion effect. To further improve the expressive ability of the factor matrices, the present invention introduces singular value decomposition (SVD) to perform modal expansion and feature extraction on the experimental tensor \(X\) and the core tensor \(G\).

[0072] First, expand the observed tensor \(X\) and the core tensor \(G\) along each order of the mode respectively, denoted as:

[0073] \(X\) t =\(\Gamma(X,t)\), \(G\) t =\(\Gamma(G,t)\), \(t\in[1:4]\)

[0074] where the functions \(\Gamma(X,t)\) and \(\Gamma(G,t)\) represent the \(t\)-th order modal expansion of \(X\) and \(G\). The matrix is used to convert the high-order tensor into a two-dimensional matrix for matrix decomposition, is used to extract the global correlation features of the core tensor at each order. Taking the first-order expansion matrix \(X_1\in R\) 288×(28×13×13) as an example, it represents the time dimension expansion matrix.

[0075] Subsequently, perform SVD decomposition on the expansion matrices \(X\) t and \(G\) t :

[0076]

[0077] where, is the left singular vector matrix of the matrix \(X\) t , with the dimension of \(I\) t \(\times I\) t . Its column vectors correspond to the left singular vectors of \(X\) t , and are used to capture the local spatio-temporal patterns of the original data at the \(t\)-th order mode. is the diagonal singular value matrix, with the dimension of \(I\) t \(\times\Pi\) k≠t \(I\) k . The diagonal elements are arranged in descending order, reflecting the importance of each singular vector. The vectors corresponding to the larger singular values contain more main features. is the right singular vector matrix of the matrix \(X\) t , with the dimension of \(\Pi\) k≠t \(I\) k \(\times\Pi\) k≠t \(I\) k . Its column vectors correspond to \(X\)t The right singular vectors assist in completing the orthogonal decomposition of the matrix. is the matrix G t The left singular vector matrix of, with dimension K t ×K t . Its column vectors are the global correlation patterns of the core tensor at the t-th order, used to guide the global fusion of local features. is the diagonal singular value matrix, with dimension K t ×Π k≠t K k . The diagonal elements are arranged in descending order, reflecting the global feature importance of the core tensor at each order. is the matrix G t The right singular vector matrix of, with dimension Π k≠t K k ×Π k≠t K k . Its column vectors are the right singular vectors of the core tensor at the t-th order, participating in the orthogonal decomposition of global features.

[0078] In singular value decomposition, the magnitude of each singular value directly reflects the information energy carried by its corresponding singular vector. Generally speaking, the sum of the first 10% of the singular values usually occupies the vast majority of the sum of all singular values, which means that these singular vectors have more features. The present invention selects the first k t vectors corresponding to the largest singular values in, to construct the local feature matrix At the same time, select the first k t vectors in, to construct the global feature matrix

[0079] Construct the fused feature factor matrix through matrix multiplication:

[0080]

[0081] Use the core tensor G and the fused factor matrices R1, R2, R3, R4 to reconstruct the complete tensor, and the calculation formula is:

[0082]

[0083] is the tensor reconstructed by C-Tucker, integrating the local and global features of each order, realizing the simultaneous repair of discrete missing and continuous missing data, and ensuring the comprehensiveness and accuracy of traffic data completion.

[0084] To further ensure the low-rank structure, a nuclear norm constraint can be introduced to obtain the objective function:

[0085]

[0086] Among them, is the Frobenius norm, W is the missing mask tensor, the parameter λ is used to control the weight of the low-rank constraint, and ||X t || represents the nuclear norm of the matrix after the tensor X is unfolded along the t-th order mode.

[0087] To better measure the present invention, the present invention uses the PeMS highway dataset for experiments, which consists of traffic flow data at 13 monitoring points from location A to location M. And the root mean square error (RMSE) is used as a quantization index for data imputation error. The smaller the RMSE value, the smaller the difference between the observed value and the estimated value, and thus the better the imputation effect. The coefficient of determination (R-Square) is used to measure the correlation or consistency between the estimated value and the true observed value when the model recovers the missing data. The value of R-Square ranges from 0 to 1, and the closer it is to 1, the better the fitting effect of the model on the observed data. The comparison methods selected are the high-precision low-rank tensor completion model (HaLRTC) and the low-rank truncated nuclear norm tensor completion model (LRTC-TNN).

[0088] Example 1

[0089] To compare the imputation performance of the present invention and the comparison methods in the case where both discrete and continuous random missing data exist. The discrete random missing rate in the experimental data is fixed at 1%, while for the continuous random missing data, different sizes are set in space and time. Among them, the missing data appears at the detection points {G}, {F, G, H}, and {E, F, G, H, I} respectively, and the number of consecutive random missing weeks at each detection point is set from 1 to 15 weeks. Figure 4 It shows that when the number of detection points is fixed, the RMSE of HaLRTC shows an upward trend as the number of consecutive missing weeks increases, indicating that the imputation effect becomes worse. While the RMSE curve of the LRTC-TNN method shows a downward trend, and the RMSE curve of the present invention is relatively stable, and the overall performance is better.

[0090] Example 2

[0091] From Figure 5 the R-Square index in it can be seen that when the number of consecutive missing weeks is small, the fitting effect of HaLRTC is better; while when the number of consecutive missing weeks increases, the present invention can maintain a high R-Square value, indicating that it has higher robustness in the imputation of continuous missing data. Generally speaking, the present invention shows a relatively stable and better imputation effect than other methods for different numbers of consecutive missing data weeks.

[0092] Example 3

[0093] To further verify the applicability of the method of the present invention under different missing rates, it is carried out at the detection points {E, F, G, H, I}, with the fixed discrete random missing rate being 1%, while the continuous missing rates are respectively set from 5% to 80%.

[0094] The experimental results are as Figure 6 shown. The method of the present invention performs best when the continuous missing rate is greater than 5%. Its RMSE is stable at about 23, and the R-Square value remains at about 0.75. In addition, with the further increase of the missing rate, the filling effect of the present invention is slightly improved, indicating its high robustness.

Claims

1. A traffic missing data completion method based on C-Tucker fourth-order spatio-temporal tensor representation, characterized in that It includes the following steps: Step 1: Based on the spatio-temporal correlation of traffic flow data, construct a fourth-order spatio-temporal tensor model including four dimensions of time, day, cycle, and space, and initialize the observation tensor X; Step 2: Based on the fourth-order spatio-temporal tensor model, perform Tucker decomposition on the observation tensor X to calculate the core tensor G; Step 3: Calculate the nth-order mode expansion of the observed tensor X and the core tensor G, extract the set of eigenvectors using singular value decomposition (SVD), and select the eigenvectors with the largest eigenvalues according to the dimension of each order to form the factor matrix N t and the core matrix M t ; Step 4. Construct a factor matrix N based on the eigenvectors t and a core matrix M t to construct a fusion matrix R t ; Step 5: Reconstruct the observation tensor X through the core tensor G and the fusion matrix R t to achieve the recovery of traffic data.

2. The traffic missing data completion method based on C-Tucker fourth-order spatio-temporal tensor representation according to claim 1, wherein Step 1 specifically is: Based on the spatio-temporal correlation of traffic flow data, construct a fourth-order spatio-temporal tensor model It includes the following dimensions: Time, I1 = 288: With a 5-minute time interval, each day is divided into 288 time slices to capture the high and low fluctuations of the intraday traffic; Day, I2 = 28: Taking 4 weeks (i.e., 28 days) as a single cycle to model the traffic changes and their correlations between different days; Cycle, I3 = 13: The 52 weeks of the whole year are divided into 13 cycles, and each cycle contains 4 weeks of data to explicitly encode the long-term patterns across months and quarters; Location, I4 = 13: Using 13 spatially adjacent detection points to improve the robustness during completion through proximity constraints; The complete traffic flow data is represented as a fourth-order tensor A ∈ R 288×28×13×13 , where each dimension represents time, day, period, and location in sequence, and it contains 1,362,816 elements in total.

3. A traffic missing data completion method based on C-Tucker fourth-order spatio-temporal tensor representation according to claim 1, characterized in that Step 2 specifically is: construct a binary mask tensor W i,j,k,l ∈ {0, 1} 288×28×13×13 , and the expression is as follows: Then, obtain the observation tensor X through the Hadamard product operation: Among them, represents the symbol of Hadamard product operation. A is the complete data set. In the operation result, the missing data is disposed as 0, and the known data is retained, so as to separate the missing part and the observed part in the traffic data; Perform Tucker decomposition on the observed tensor X to obtain the core tensor G and factor matrices U of each order t , t ∈ [1:4]: X ≈ G × 1U1 × 2U2 × 3U3 × 4U4 Among them, the core tensor stores the global correlation features across dimensions and satisfies K t <I t , t ∈ [1:4]; the factor matrices respectively represent the local patterns of each dimension.

4. A traffic missing data completion method based on C-Tucker fourth-order spatio-temporal tensor representation according to claim 1, characterized in that, Step 3 is specifically as follows: First, expand the observation tensor X and the core tensor G along each order of the mode respectively, denoted as: X t = Γ(X, t), G t = Γ(G, t), t ∈ [1:4] Among them, the functions Γ(X, t) and Γ(G, t) represent the t-order modal expansions of X and G; the matrix is used to convert a high-order tensor into a two-dimensional matrix for matrix decomposition, and is used to extract the global correlation features of the core tensor at each order; taking the first-order expansion matrix X1 ∈ R 288×(28×13×13) as an example, it represents the time-dimensional expansion matrix; For the unfolded matrix X t and G t perform SVD decomposition: Among them, is the left singular vector matrix of matrix X t with dimension I t ×I t ; its column vectors correspond to the left singular vectors of X t and are used to capture the local spatio-temporal patterns of the original data at the t-th order mode; is the diagonal singular value matrix with dimension I t ×Π k≠t I k ; the diagonal elements are arranged in descending order, reflecting the importance of each singular vector; is the right singular vector matrix of matrix X t with dimension Π k≠t I k ×Π k≠t I k ; its column vectors correspond to the right singular vectors of X t and assist in completing the orthogonal decomposition of the matrix; is the left singular vector matrix of matrix G t with dimension K t ×K t ; its column vectors are the global correlation patterns of the core tensor at the t-th order, used to guide the global fusion of local features; is the diagonal singular value matrix with dimension K t ×Π k≠t K k ; the diagonal elements are arranged in descending order, reflecting the importance of the global features of the core tensor at each order; is the right singular vector matrix of matrix G t with dimension Π k≠t K k ×Π k≠t K k ; its column vectors are the right singular vectors of the core tensor at the t-th order and participate in the orthogonal decomposition of the global features; In singular value decomposition, the magnitude of each singular value reflects the information energy carried by its corresponding singular vector. Select the first k t vectors corresponding to the largest singular values to construct a local feature matrix Meanwhile, select the first k t vectors to construct a global feature matrix 5. A traffic missing data completion method based on C-Tucker fourth-order spatio-temporal tensor representation according to claim 4, characterized in that Step 4 is specifically as follows: Construct the fused feature factor matrix through matrix multiplication: Among them, N t represents the local feature matrix, and M t represents the global feature matrix.

6. A method for completing missing traffic data based on C-Tucker fourth-order spatio-temporal tensor representation according to claim 5, characterized in that, Step 5 is specifically as follows: Use the core tensor G and the fused factor matrices R1, R2, R3, R4 to reconstruct the complete tensor, and the calculation formula is: is a tensor reconstructed by C-Tucker, integrating local and global features of each order of modality to simultaneously repair discrete missing and continuous missing data; Introduce the nuclear norm constraint to further ensure the low-rank structure and obtain the objective function: Among them, is the Frobenius norm, W is the missing mask tensor, and the parameter λ is used to control the weight of the low-rank constraint. ||X t || represents the nuclear norm of the matrix after the observed tensor X is unfolded along the t-th order mode.

7. A computer device, comprising a memory, a processor, and a computer program stored on the memory, characterized in that, The processor executes the computer program to implement the steps of the method described in claim 1.

8. A computer-readable storage medium having computer programs / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, the steps of the method described in claim 1 are implemented.

9. A computer program product, comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, the steps of the method described in claim 1 are implemented.

Images