A cross-travel platform supply and demand forecasting method based on graph federated learning

Abstract

The application discloses a cross-travel platform graph federated learning prediction method, and belongs to the field of federated computing. The center server is responsible for coordinating various travel platforms, operator decoupling, model aggregation and model updating; the travel platform side mainly cooperates with the server to perform local space-time graph modeling and local space-time feature extraction. Each travel platform first divides the service area into different grid areas; then, according to historical data, the supply and demand data of different areas are calculated, and the historical supply and demand in the area are taken as prediction features; subsequently, the association relationship between different areas is modeled into a graph federation, and cross-platform feature aggregation is performed under the coordination of the center server, so that cross-platform space-time supply and demand prediction is realized. The application reduces the computing overhead and takes into account privacy and efficiency.

Description

Technical Field

[0001] The present invention belongs to the field of federated computing and relates to a supply and demand forecasting method based on graph federated learning, and specifically to a cross-travel platform supply and demand forecasting method based on graph federated learning. Background Art

[0002] The technical fields related to the present invention mainly include two: a supply and demand forecasting method for large-scale shared travel and a privacy protection method for federated spatiotemporal graph forecasting.

[0003] Most previous supply and demand forecasting methods for shared mobility have focused solely on the domain of a single platform. These methods typically model spatial regions as graph nodes and then combine them with time series models for forecasting. These methods are limited by the data from a single platform and are unable to leverage data from various mobility platforms to form a global perspective. This can lead to limited accuracy in supply and demand forecasts, often known as "prediction bias." Recent approaches have leveraged federated learning to train machine learning models using data from across multiple institutions, often achieving significant improvements over models trained on single-institution data.

[0004] Privacy-preserving technologies for federated spatiotemporal graph prediction fall into two main categories: graph embedding-based technologies and homomorphic encryption-based technologies. In graph embedding-based technologies, each federated member represents its private spatiotemporal graph as an embedding vector, which is assumed to effectively protect privacy. After obtaining the embedding vectors, the federated central server can directly apply the same spatiotemporal graph prediction method as in a single-platform scenario. However, the privacy of graph embeddings is lacking in both theory and practice, posing privacy risks of raw data leakage. In homomorphic encryption-based technologies, some work allows federated members to share a common public key to encrypt the various intermediate results transmitted during federated prediction. In this scenario, traditional spatiotemporal graph prediction methods for a single mobility platform can be simply extended to a cross-platform federated prediction scenario using encrypted data. This allows for supply and demand prediction to be performed using historical supply and demand data from different regions across various shared mobility platforms while maintaining data privacy. However, this technology incurs high communication and computational overhead, making it difficult to achieve timely supply and demand prediction across shared mobility platforms.

[0005] In summary, existing methods still lack a federated supply and demand forecasting method for cross-shared travel platforms. Moreover, for commonly used spatiotemporal forecasting technologies, especially the frequently used spatiotemporal graph forecasting method, there is still a lack of a federal privacy protection framework that takes into account both data privacy security and the timeliness of supply and demand forecasting. Summary of the Invention

[0006] By analyzing and comparing existing technologies, the present invention designs a cross-travel platform supply and demand prediction method based on graph federated learning, which allows multiple shared travel platforms to aggregate the supply and demand characteristics of each party in different time and space regions under the data privacy requirement of "raw data does not leave the domain, data is available but not visible", avoiding the "prediction bias" caused by being limited by the data of a single platform, and realizing high-efficiency and accurate prediction of global driver and passenger supply and demand.

[0007] A cross-mobility platform supply and demand forecasting method based on graph federated learning. The specific steps are as follows:

[0008] Step 1: Divide the data of m travel platforms into temporal and spatial supply and demand areas, and build a spatiotemporal graph federation model;

[0009] In the spatial dimension, each travel platform is divided into grid areas of different sizes according to its geographical location; the grid areas are used as nodes of the space-time graph, and the connecting edges of different grid areas are constructed through space-time learning methods.

[0010] In the time dimension, the maximum sampling frequency of each travel platform is selected as the federated time window size, and the time window is processed as the historical supply and demand forecast feature of the sliding time window;

[0011] The maximum sampling frequency of the travel platform o is: l = max{l o},o∈O, where set O represents the set of travel platforms, and lo represents the GPS sampling frequency of the capacity of travel platform o;

[0012] Finally, through time and space division, a local time and space map of supply and demand forecasts for each travel platform is constructed Indicates that travel platform o is in time window t T The local space-time graph of each travel platform G o Together they form a spatiotemporal graph federation model G = {G1,…,G m}.

[0013] Step 2: Extract local spatiotemporal features of each travel platform to obtain spatial correlation matrix and temporal features;

[0014] In the time dimension, the time series model is used to process the time series features of the data of each travel platform in the local time-space graph, and the time window tT of the travel platform o is converted into his : The supply and demand characteristics within t are expressed as The time characteristics of m travel platforms are expressed as

[0015] Time series models include gated recurrent neural networks, long short-term memory neural networks, etc.

[0016] In the spatial dimension, the local spatio-temporal graph federated correlation weight E is learned based on historical data, E = [E1,..., E m ], where E o represents the correlation of the grid area of the ride platform o.

[0017] Step three, the center server uses the local spatio-temporal characteristics of each ride platform to decouple the spatio-temporal prediction operator, and uses it for local calculation of each ride platform;

[0018] The spatio-temporal prediction operator decoupling refers to that the center server decouples the distributed graph convolution operator into an operator that can be executed by each platform independently, and then each ride platform completes local calculation. The specific description is as follows:

[0019] Solve the optimization problem to obtain the quadratic function approximation of the ReLU function:

[0020] The optimization problem is:

[0021]

[0022] When p = 1 is selected, based on the optimization target of the L1 loss function, the approximate activation function of ReLU(x) is obtained as: ReLU2(x) = 0.867x 2 + 0.499x + 0.0395.

[0023] After adopting the ReLU2(x) activation function, the calculation of the auxiliary spatio-temporal graph convolution operator A = ReLU2(E·E T )h t is decoupled, and the specific process is as follows:

[0024] The activation function ReLU2(x·y) is equivalent to the transformation ReLU 2,2 (x)·y 2 + ReLU 2,1 (x)y + ReLU 2,0 (x), where ReLU 2,k (x),k = 0,1,2 represents the kth term of the ReLU2(x) activation function. Therefore, the spatio-temporal graph convolution operator ReLU(EE T )h t is decoupled and transformed into the following form:

[0025]

[0026] where, represents the tensor product of the matrix E o itself, and E o ·E p represents the inner product of the matrix.

[0027] After decoupling transformation, each travel platform only needs local calculation The internal correlation matrix tensor product is sufficient.

[0028] Step 4: Aggregate the local calculation results of each travel platform and input them into the global spatiotemporal graph federation model to perform global spatiotemporal supply and demand forecasting;

[0029] The various calculation results are aggregated through a secure aggregation operation based on random masks and input into the global model as acceptable historical features of the sliding window in the global spatiotemporal graph prediction to complete the global spatiotemporal supply and demand prediction.

[0030] Aggregation is done based on a random mask secure summation method. The calculation method is as follows:

[0031] For m travel platforms, each platform holds a privacy value v o ; Secure summation over random masks allows each party to generate a mask:

[0032]

[0033] Among them, sd o With sd o' are the random number seeds of platforms o and o' respectively, is a pseudo-random generator function. Under this mask, the sum of the masked privacy values ​​of all parties is equal to the sum of the original privacy values, that is,

[0034] Therefore, the quadratic and linear terms of ReLU2 in step 3 can both be securely summed using random masks to complete the calculation of the correlation matrix across travel platforms for time window t.

[0035] Then, As historical features, it is input into the global spatiotemporal graph federation model to predict the future time window t+1 to t+T fut The travel supply and demand situation in different regions within the country.

[0036] In step 5, the central server updates the global spatiotemporal graph federation model based on the global spatiotemporal supply and demand forecast results, and distributes the global spatiotemporal supply and demand forecast and the new global spatiotemporal graph federation model to each travel platform. Each travel platform updates its local spatiotemporal graph prediction model and returns to step 1 to iterate local and global federation calculations until the supply and demand model of the spatiotemporal graph federation learning converges.

[0037] Through the loss function Update global model parameters Obtain a new global spatiotemporal graph federation model G θ′ .

[0038] Step six, according to the global prediction results and local historical data, the supply and demand of each travel platform in different regions are predicted.

[0039] In summary, the advantages of the present application are as follows:

[0040] (1) The method establishes a spatio-temporal graph federation as a model for cross-travel platform supply and demand prediction. Different grid regions are divided as nodes of the spatio-temporal graph in space, and a historical supply and demand data driven inter-regional correlation matrix is constructed by machine learning method; in time, the historical features and dynamic correlation matrix of each platform are received by using the sliding window method.

[0041] (2) The method designs a quadratic function approximation calculation form of ReLU(x) activation function in spatio-temporal graph convolution, and innovatively uses L1 loss function as the optimization target for measuring the approximation effect, and finally obtains the quadratic approximation activation function ReLU2(x) near the mean of 0.

[0042] (3) With the help of the approximate activation function ReLU2(x), the binary function ReLU(xy) is decoupled, allowing each platform to decouple the correlation matrix into a form that can be used for low-computing-cost mask-safe aggregation, avoiding the high computing cost of traditional homomorphic encryption methods, and balancing privacy and efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 The workflow diagram of the cross-travel platform supply and demand prediction method based on graph federated learning of the present application;

[0044] Figure 2 The experimental results of different methods for sharing travel supply and demand prediction in the embodiment of the present application;

[0045] Figure 3 The experimental results of the quadratic activation function approximation error used in the present application. DETAILED DESCRIPTION

[0046] The present application will be further described in detail below in combination with the drawings and examples.

[0047] The spatio-temporal supply and demand prediction of different regions can use the commonly used sliding time window mode, that is, each platform uses the historical data in the past T his time window as input to predict the supply and demand of the future T fut time window. Assuming that the current time is t, the prediction task can be represented as:

[0048]

[0049] G θA spatio-temporal graph model representing the supply-demand prediction, and theta represents the learnable parameters of the model.

[0050] For cross-platform sharing travel, the spatio-temporal characteristics are scattered by each platform, that is, X = {X1,..., X m Due to data privacy constraints, each platform cannot directly share its historical characteristics out of domain; therefore, a privacy-preserving method is needed to train and infer the spatio-temporal graph model while utilizing data from each party.

[0051] Considering the spatio-temporal data characteristics of cross-sharing travel platforms, the present application designs a cross-travel platform supply-demand prediction method based on spatio-temporal graph federation, which takes into account both computational efficiency and data privacy.

[0052] The implementation platform of the present application is a supply-demand prediction platform for cross-sharing travel, and hereinafter, a different regional spatio-temporal prediction method is introduced using a ride-hailing platform as an example. The system framework of the present application is shown in Figure 1 The central server is responsible for coordinating various travel platforms, operator decoupling, model aggregation, and model updating, etc.; while the travel platform side mainly cooperates with the server to perform local spatio-temporal graph modeling and local spatio-temporal feature extraction. Each travel platform first divides the service area into different grid regions (such as hexagons, squares, etc.); then, according to historical data, the supply-demand data of different regions is calculated, and the historical supply-demand in the region is taken as the prediction feature; subsequently, the association between different regions can be modeled into a graph federation, and cross-platform feature aggregation is performed under the coordination of the central server to realize cross-platform spatio-temporal supply-demand prediction. Specifically, it can be divided into the following steps:

[0053] Step 1: Spatio-temporal supply-demand region division:

[0054] Unlike most previous prediction techniques that construct road networks (intersections, roads) as spatio-temporal graphs, the present application adopts a new spatio-temporal graph federation construction method in time and space. Specifically:

[0055] In the spatial dimension, each travel platform loads historical capacity supply and user travel data for clustering analysis and divides them into different size grid regions (such as square grids with a side length of 5 km, hexagonal grids, etc.), and corresponding historical supply-demand data of different regions is calculated, and then different spatial regions can be constructed as nodes to build a federated spatio-temporal graph. The connection edges of different grid regions are constructed by spatio-temporal learning method. Since the user initiates the trip in the shared travel, the location is often not on the road, so using grid as the basic unit of supply-demand prediction can better fit the supply-demand prediction task of shared travel.

[0056] In the time dimension, in order to avoid the problem of "ghost car" caused by inconsistent sampling frequency of each platform (part of the capacity may suddenly appear in a certain time window due to inappropriate sampling frequency), the maximum sampling frequency of the travel platform is selected as the federal time window size, l = max {l o}, o e O, where the set O represents the set of travel platforms, and lo represents the capacity GPS sampling frequency of the travel platform o;

[0057] Finally, different spatial regions are taken as nodes of the graph, and time regions are taken as node features of the graph, to construct a spatio-temporal graph federation model. The local spatio-temporal graph G The local spatio-temporal graph G o of each travel platform jointly constitutes a spatio-temporal graph federation model G = {G1,..., G m} for demand prediction of m travel platforms.

[0058] Step 2: Local spatio-temporal feature extraction:

[0059] In the time dimension, the commonly used threshold recurrent neural network or long short-term memory neural network is adopted to process the time sequence features in the local spatio-temporal graph, and the demand and supply features of the travel platform o in the time window t-T his are represented as The time features of m travel platforms are represented as

[0060] In the spatial dimension, the local linkage graph federation correlation weight E = [E1,..., E m ] is learned based on historical data, where E o represents the correlation of the grid region of the travel platform o.

[0061] Step 3: Spatio-temporal prediction operator decoupling:

[0062] The center server assisting the demand prediction of each platform needs to jointly use the cross-platform time features and spatial features to complete the cross-platform demand prediction, which decouples the distributed graph convolution operator into an operator that can be executed by each platform independently; then, each shared travel platform completes local calculation.

[0063] To realize the spatio-temporal demand prediction across platforms, the spatio-temporal correlation matrix across platforms (i.e. the influence of the travel platform o on another travel platform p) needs to be calculated first, which depends on the activation operator ReLU (EE T )h t However, the platform o only holds its own spatial correlation matrix E o and time features To solve the above problem, the traditional method usually adopts the same set of homomorphic encryption public keys for E o and Encryption is performed, which requires pairwise communication between each platform However, this method has a huge computational overhead, and the prediction time required will exceed the size of the time window, making it impossible to apply in practice.

[0064] To solve the above problems, the application proposes an activation operator approximation method based on a quadratic function: each travel platform calculates the ReLU activation function under the quadratic function approximation, and decouples the intermediate results of the cross-domain correlation matrix required by each platform with the help of the ReLU activation function; then, a random number seed that can generate a mask is generated, and a random mask is added to the local intermediate results under the protocol, providing preparation for model aggregation.

[0065] The quadratic function approximation of the ReLU function is obtained by solving the optimization problem.

[0066] The optimization problem is:

[0067]

[0068] Unlike the L2 loss function in the traditional method, the application innovatively selects an optimization target based on the L1 loss function (i.e., p = 1), which can obtain a better approximation effect when the average value of the calculated number x is near 0.

[0069] By solving the above optimization target, the approximate activation function ReLU2(x) of ReLU(x) can be obtained as follows: ReLU2(x) = 0.867x 2 + 0.499x + 0.0395.

[0070] After adopting the ReLU2(x) activation function, the calculation of the spatio-temporal graph convolution operator A = ReLU2(E·E T )h t can be decoupled. ReLU2(x·y) is equivalent to ReLU 2,2 (x)·y 2 + ReLU 2,1 (x)y + ReLU 2,0 (x), where ReLU2,k(x) represents the kth term of the ReLU2(x) activation function. Therefore, the spatio-temporal graph convolution operator ReLU(EE T )h t can be decoupled and transformed as follows:

[0071]

[0072] where, represents the tensor product of the matrix E o itself, and E o ·E p represents the inner product of the matrix.

[0073] After decoupling transformation, each shared travel platform only needs to calculate locally The internal correlation matrix tensor product is sufficient.

[0074] Step 4: Aggregate the spatiotemporal prediction results. Aggregate the intermediate results of the cross-domain correlation matrix of each travel platform to obtain the spatiotemporal correlation matrix under the global supply and demand, and complete the cross-platform supply and demand prediction based on this matrix.

[0075] After each shared travel platform executes the decoupled operators locally, it aggregates the calculation results through a safe aggregation operation based on random masks, and inputs them into the global model as acceptable historical features of the sliding window in the global spatiotemporal graph prediction to complete the global spatiotemporal supply and demand prediction.

[0076] The aggregation of spatiotemporal prediction results is completed based on the random mask secure summation method. The calculation method is as follows:

[0077] For m travel platforms, each platform holds a privacy value v o ; Secure summation over random masks allows parties to generate masks:

[0078]

[0079] Among them, sd o With sd o' are the random number seeds of platforms o and o' respectively, is a pseudo-random generator function. Under this mask, the sum of the masked privacy values ​​of all parties is equal to the sum of the original privacy values, that is,

[0080] Therefore, the quadratic and linear terms of ReLU2 in step 3 can both be securely summed using random masks to complete the calculation of the correlation matrix across travel platforms for time window t.

[0081] Then, you can As historical features, it is input into the global spatiotemporal graph federation model to predict the future time window t+1 to t+T fut Travel supply and demand forecasts for different regions within the country.

[0082] Step 5: Federated Model Update Aggregation:

[0083] The accuracy of the prediction results is calculated based on the root mean square error. where |G o | represents the spatial area of ​​each platform, y g and They represent the actual supply and demand and the predicted supply and demand of the spatial region g in the platform respectively;

[0084] Through the loss function updating the global model parameters to obtain a new global spatio-temporal graph federated model G θ′ and distributing the global spatio-temporal supply-demand prediction result and the new spatio-temporal graph federated model to each platform, and each platform updating a local spatio-temporal graph prediction model according to the federated model

[0085] Finally, steps 1 to 5 are repeated, and local calculation and global federated calculation iterations are performed until the supply-demand model of the spatio-temporal graph federated learning converges.

[0086] The embodiment is verified on a real data set of 4 taxi companies in Beijing, and a schematic diagram is as shown in Figure 2 It can be seen that the cross-platform supply-demand prediction error is significantly lower than the single-platform prediction after using the method, and is close to the ideal prediction error when the data is directly integrated out of the domain without considering the data privacy constraint. Figure 3 The approximation effect diagram of the ReLU2(x) function proposed in the embodiment can be seen, and it can be seen that the ReLU2(x) function better approximates the ReLU(x) activation function in the range of [-0.5, 0.5].

Claims

1. A cross-travel platform supply and demand forecasting method based on graph federated learning, characterized by: The specific steps are as follows: Step 1: Divide the data of m travel platforms into temporal and spatial supply and demand areas, and build a spatiotemporal graph federation model; In the spatial dimension, each travel platform is divided into grid areas of different sizes according to its geographical location. The grid areas are used as nodes in the space-time graph, and the connection edges between different grid areas are constructed through the space-time learning method. In the time dimension, the maximum sampling frequency of each travel platform is selected as the federated time window size, and the time window is processed as the historical supply and demand forecast feature of the sliding time window; Finally, through time and space division, a local time and space map of supply and demand forecasts for each travel platform is constructed Indicates that travel platform o is in time window t T The local space-time graph of each travel platform G o Together they form a spatiotemporal graph federation model G = {G1,…,G m }; Step 2: Extract local spatiotemporal features of each travel platform to obtain spatial correlation matrix and temporal features; In the time dimension, the time series model is used to process the time series features of the data of each travel platform in the local time-space graph, and the time window tT of the travel platform o is converted into his : The supply and demand characteristics within t are expressed as The time characteristics of m travel platforms are expressed as In the spatial dimension, the local spatiotemporal graph federation association weights are learned based on historical data, E=[E1,…,E m ], where E o Represents the relevance of the grid area of ​​travel platform o; Step 3: The central server uses the local spatiotemporal features of each travel platform to decouple the spatiotemporal prediction operators and use them for local calculations on each travel platform. The decoupling of spatiotemporal prediction operators refers to the central server decoupling the distributed graph convolution operators into operators that can be executed independently by each platform, and then each travel platform completes the local calculation; Solve the optimization problem and obtain a quadratic function approximation to the ReLU function: The optimization problem is: When p=1 is selected, based on the optimization objective of the L1 loss function, the approximate activation function of ReLU(x) is solved as: ReLU2(x)=0.867x 2 +0.499x+0.0395; After using the ReLU2(x) activation function, the auxiliary spatiotemporal graph convolution operator A=ReLU2(E·E T )h t The calculation decoupling is as follows: Transform the activation function ReLU2(x·y) into ReLU equivalently 2,2 (x)·y 2 +ReLU 2,1 (x)y+ReLU 2,0 (x), where ReLU 2,k (x), k=0,1,2 represents the kth term of the ReLU2(x) activation function; therefore, the spatiotemporal graph convolution operator ReLU(EE T )h t The decoupling transformation is as follows: in, Represents the matrix E o The tensor product with itself, E o ·E p represents the inner product of a matrix; After decoupling transformation, each travel platform only needs local calculation The tensor product of the internal correlation matrix is ​​sufficient; Step 4: Aggregate the local calculation results of each travel platform and input them into the global spatiotemporal graph federation model to perform global spatiotemporal supply and demand forecasting; The individual calculation results are aggregated through a secure aggregation operation based on random masks and input into the global model as acceptable historical features of the sliding window in the global spatiotemporal graph prediction to complete the global spatiotemporal supply and demand forecast; The quadratic and linear terms of ReLU2 in step 3 can be safely summed using random masks to complete the calculation of the correlation matrix across travel platforms for time window t; Then, As historical features, it is input into the global spatiotemporal graph federation model to predict the future time window t+1 to t+T fut Travel supply and demand in different regions within the country; Step 5: The central server updates the global spatiotemporal graph federation model based on the global spatiotemporal supply and demand forecast results, and distributes the global spatiotemporal supply and demand forecast and the new global spatiotemporal graph federation model to each travel platform. Each travel platform updates its local spatiotemporal graph prediction model and returns to step 1 to iterate local and global federation calculations until the supply and demand model of the spatiotemporal graph federation learning converges. Step 6: Predict the capacity supply and user demand of each travel platform in different regions based on the global prediction results and local historical data.

2. The cross-travel platform supply and demand forecasting method based on graph federated learning according to claim 1 is characterized in that: The maximum sampling frequency of the travel platform o is: l = max{l o },o∈O, where set O represents the set of travel platforms, l o Indicates the GPS sampling frequency of the transportation capacity of travel platform o.

3. The cross-travel platform supply and demand forecasting method based on graph federated learning according to claim 1 is characterized in that: Aggregation is done based on a random mask secure summation method. The calculation method is as follows: For m travel platforms, each platform holds a privacy value v o ; Secure summation over random masks allows each party to generate a mask: Among them, sd o With sd o' are the random number seeds of platforms o and o' respectively, is a pseudo-random generation function; under this mask, the sum of the masked privacy values ​​of all parties is allowed to be equal to the sum of the original privacy values, that is, 4. The cross-travel platform supply and demand forecasting method based on graph federated learning according to claim 1 is characterized in that: The specific steps for updating the global spatiotemporal graph federation model are: Through the loss function Update global model parameters Obtain a new global spatiotemporal graph federation model G θ′ .

Images