Group Division Method and System Based on Dynamic Hypergraph Clustering of Individual Movement Patterns

Abstract

The present invention relates to the field of transportation information engineering, and discloses a method and system for group division based on dynamic hypergraph clustering of individual movement patterns, including: constructing an individual movement pattern feature matrix for each time slice; constructing an individual movement pattern hypergraph for each time slice to model the complex high-order association relationship of "many-to-many" between individual movement patterns in the whole time period; proposing a dynamic hypergraph clustering model based on movement patterns to perform category division of individual associated groups; applying the proposed model to a series of empirical data sets for training and testing, and proving that it can achieve good group division performance superior to the baseline model.

Description

Technical Field

[0001] The present invention relates to the technical field of transportation information engineering, and particularly to a method and system for group division based on dynamic hypergraph clustering of individual movement patterns. Background Art

[0002] Intelligent transportation, also known as smart transportation, refers to the effective penetration and integration of the traditional transportation industry and the Internet by means of technologies and concepts such as mobile Internet, cloud computing, big data, and the Internet of Things. On this basis, real-time sensing and analysis technologies of driving behavior can also be integrated to achieve multi-mode and multi-standard dynamic navigation for public travel and improve travel efficiency.

[0003] Dividing the categories of travel groups is an important strategy for realizing intelligent transportation, which helps to better understand the needs and preferences of different user groups, so as to provide more personalized and customized services, as well as more refined proactive management.

[0004] In related technologies, generally, the division is carried out from two perspectives of time and space, and then further inferences about group categories are made based on time and space. However, the classification methods in related technologies have a high misjudgment rate. For example, due to chance or other reasons, the travel time and travel space of an individual conform to a certain group and are classified into the corresponding group, but in fact, the individual belongs to another group. In this way, the division of group categories is not accurate enough, which may further affect activities such as the formulation of corresponding policy opinions and the development of marketing. Summary of the Invention

[0005] The present invention provides a method and system for group division based on dynamic hypergraph clustering of individual movement patterns to solve the defect that the division of group categories in related technologies is not accurate enough and the misjudgment rate is high. In the solution of the present application, the travel records of individuals are clustered by constructing a hypergraph, which improves the accuracy of group category division.

[0006] The present invention provides a method for group division based on dynamic hypergraph clustering of individual movement patterns, including: obtaining a set of historical travel records of N individuals, cutting the set in the time dimension on a daily basis to obtain a set of historical travel records of N individuals under T consecutive time slices, extracting D-dimensional travel features for each individual, and constructing a movement pattern feature matrix of N individuals;

[0007] For any time slice, regarding each individual as a node to obtain N nodes; regarding the neighboring groups of the individuals as hyperedges of the hypergraph to obtain N hyperedge sets, and constructing an individual movement pattern hypergraph under each time slice;

[0008] Construct a dynamic hypergraph clustering model based on movement patterns, that is, rely on the "encoding-decoding" structure of Transformer to learn the deep expression of individual movement patterns; rely on the hypergraph convolutional network model to learn the deep expression of the movement patterns of neighborhood groups in the local space; use the attention mechanism to fuse the above deep expressions of individual and neighborhood group movement patterns, and use the bidirectional hypergraph convolutional network to capture the change value of the hypergraph topology under the cross-action of forward time propagation and backward time propagation, and dynamically update the hypergraph structure; complete the training and testing of the model;

[0009] Input the historical travel record set of M individuals into the dynamic hypergraph clustering model of movement patterns for instance verification of individual associated group category division.

[0010] According to the group division method based on individual movement pattern dynamic hypergraph clustering provided by the present invention, extract D-dimensional travel features for each individual, and construct a movement pattern feature matrix of N individuals, including:

[0011] At any time slice τ, extract D-dimensional travel features for each individual, and construct a movement pattern feature matrix X of N individuals τ ∈R N×D , Obtain the individual movement pattern feature matrix X = {X1,..., X τ ,..., X T} under T consecutive time slices, where x τ,i (1 ≤ i ≤ N) is the movement pattern feature expression of the i-th individual at the τ-th time slice.

[0012] According to the group division method based on individual movement pattern dynamic hypergraph clustering provided by the present invention, the construction of the individual movement pattern hypergraph at each time slice includes:

[0013] For any time slice τ, regard the N individuals as the nodes of the hypergraph to obtain N nodes; use the Euclidean distance to measure the similarity value of the movement patterns of any two individuals; connect two adjacent nodes, and use the inverse function of the Euclidean distance as the corresponding edge weight; according to the principle that "the smaller the Euclidean distance, the more similar the movement patterns", determine the k individuals with the most similar movement patterns to form the i-th neighborhood group at the τ-th time slice

[0014] Regard the neighborhood group as a hyperedge to obtain a set of N hyperedges, form the initial division result of the individual associated group, and obtain the individual movement pattern hypergraph G τ =(V τ , E τ , X τ , W τ , Hτ ) where V τ is the set of nodes representing all individuals, and E τ is the set of hyperedges representing the movement patterns of neighborhood groups, and X τ is the set of node attributes representing the movement patterns of individuals, and W τ is the set of hyperedge weights representing the importance of neighborhood groups, and H τ is the incidence matrix representing the membership relationship between individuals and groups, and H τ = [h(v, e)] τ ∈R N×N , where h(v, e) characterizes the membership relationship between node v and hyperedge e;

[0015] The membership relationship between the node v and the hyperedge e conforms to the following formula:

[0016]

[0017] In formula (1), v ∈ e means that the node is on the hyperedge, and D τ,v = [d(v)] is the degree matrix of node v at time slice τ, which is used to describe the importance of node v in the hypergraph G τ at the current moment; d(v) is the sum of the weights of all hyperedges connected to vertex v, which is calculated by formula (2):

[0018] d(v) = ∑ e∈E ω(e)h(v, e) (2)

[0019] D τ,e = [d(e)] is the degree matrix of hyperedge e at time slice τ, which is used to describe the importance of hyperedge e in the hypergraph G τ at time slice τ, and d(e) is the sum of the number of all nodes connected by hyperedge e, which is calculated by formula (3):

[0020] d(e) = ∑ v∈V h(v, e) (3)

[0021] According to the group division method based on individual movement pattern dynamic hypergraph clustering provided by the present invention, a dynamic hypergraph clustering model based on movement patterns is constructed, including:

[0022] Using the Transformer model to learn the deep representation of individual movement patterns;

[0023] Using the hypergraph convolutional network model to learn the deep expression of neighborhood group movement patterns in the local space;

[0024] Fusing the individual and neighborhood group movement patterns for dynamic hypergraph deep clustering;

[0025] Perform the training and testing processes of the mobile pattern dynamic hypergraph clustering model to complete the dynamic division of population categories.

[0026] According to the population division method based on individual mobile pattern dynamic hypergraph clustering provided by the present invention, the use of the Transformer model to learn the deep representation of individual mobile patterns includes:

[0027] Through the working principle of the multi-head attention mechanism of the built-in encoder of the Transformer model, autonomously learn the deep expression C of the individual mobile pattern in the l-th (1 ≤ l ≤ L) convolutional layer of the individual mobile pattern feature matrix X (l) , C (l) Calculated by Equation (4):

[0028]

[0029] Among them, σ ReLU is the linear activation function. When l = 1, C (l-1) = C 0 is the individual mobile pattern feature matrix X; LN is the convolutional layer regularization operation; dropout is the operation of randomly pruning the fully connected layer; FeedForward is the fully connected network, Res (l) is the residual unit of the decoder of the Transformer model; U1 and U2 are the weight parameters to be trained by the model, b1 and b2 are the bias values; MultiHead is the multi-head attention mechanism;

[0030] Through the built-in decoder of the Transformer model, after L convolutional layers, the reconstructed Complete the decoding operation;

[0031] Minimize the average reconstruction error To obtain the depth expression form of the individual mobile pattern closest to the real situation. In the formula, ||·|| Fro is the F-norm, where the average reconstruction error is generated during the encoding and decoding processes, and the expression of the average reconstruction error is shown in formula (5):

[0032]

[0033] For any time slice τ, according to the distribution of the feature matrix X τ and the true population category to which each individual belongs, calculate the true probability distribution of individual i belonging to category j, and then obtain the individual mobile pattern category probability distribution curve GT in the real scenario at time slice τ;

[0034] According to the deep expression C of the individual mobile pattern in the L-th convolutional layer(L) The distribution situation, obtain the individual movement pattern category probability distribution curve TR predicted by the Transformer under the time slice τ;

[0035] Determine the cumulative error of the probability distribution curve GT and the probability distribution curve TR over all time slices Determine the group category division error after learning the temporal features of the Transformer, where Calculated by formula (6),

[0036]

[0037] where, GT τij and TR τij respectively refer to the probability that individual i belongs to category j in curve GT and curve TR under time slice τ, and KL is the "Kullback-Leibler divergence" distribution curve.

[0038] According to the group division method based on individual movement pattern dynamic hypergraph clustering provided by the present invention, using the hypergraph convolutional network model to learn the deep expression of the neighborhood group movement pattern in the local space, including:

[0039] Based on the hypergraph G τ and the initial division result of the individual associated group, perform hypergraph convolutional network operations to capture the topological changes in the local space of the neighborhood group. For the l-th convolutional layer, the deep expression Z of the neighborhood group movement pattern (l) is calculated by formula (7):

[0040]

[0041] where, Θ is the convolutional hyperparameter; when l = 1, Z (l-1) = Z 0 is the feature matrix X; H τ Transponse represents the transpose matrix of H τ .

[0042] According to the group division method based on individual movement pattern dynamic hypergraph clustering provided by the present invention, fuse the individual and neighborhood group movement patterns for dynamic hypergraph deep clustering, including:

[0043] Concatenate the deep expression C of the individual movement pattern (l) and the deep expression Z of the neighborhood group movement pattern (l) through formula (8) to obtain the concatenation result CZ of C (l) and Z (l) under the l-th convolutional layer, (l) CZ (l) n is CZ(l) The nth vector sequence;

[0044] Introduce the attention mechanism and obtain a new expression for the neighborhood group movement pattern in the lth convolutional layer through formula (9)

[0045]

[0046]

[0047] where q is the query vector, and α n is the attention weight of the nth element. s(CZ(l) n , q) is the attention scoring function, which is used to measure the similarity between CZ(l) n and the query vector q, and is quantitatively expressed using the dot product model;

[0048] Based on the distribution, capture the change ΔG of the hypergraph topology under the cross-action of forward time propagation and backward time propagation τ , that is, the change in the hypergraph topology that belongs to the first or second hypergraph but not both at the same time. The calculation method is shown in formula (10):

[0049] ΔG τ = difference(G 1→τ , G T→τ ) = (G 1→τ ∪ G T→τ ) - (G 1→τ ∩

[0050] G T→τ ) (10)

[0051] Based on G τ and ΔG τ , update the corresponding dynamic hypergraph G τ , which is used as the input of the hypergraph convolutional network in the (l + 1)th convolutional layer, and after performing attention fusion with the corresponding output result C (l+1) of the Transformer in the (l + 1)th layer, perform the hypergraph convolutional network operation to obtain the probability distribution result of the group categories in the (l + 1)th to Lth convolutional layers

[0052]

[0053] where W′ τ , H′ τ , D′ τ,v , D′ τ,e are the dynamic hypergraphs respectively The hyper-edge weight set, incidence matrix, node degree matrix, and hyper-edge degree matrix;

[0054] For any time slice τ, according to the deep expression Z of the individual movement pattern in the l-th convolutional layer (L) Based on the distribution, obtain the probability distribution curve HC of the individual movement pattern categories obtained by hypergraph clustering under the time slice τ, and measure the cumulative error between the probability distribution curve HC and the curve GT in all time slices Measure the group category division error of the dynamic hypergraph depth clustering that fuses individual and neighborhood group movement patterns, Calculated by Equation (12):

[0055]

[0056] Among them, HC τij Is the probability that individual i belongs to category j in the curve HC under the time slice τ.

[0057] According to the group division method based on individual movement pattern dynamic hypergraph clustering provided by the present invention, the training and testing processes of the movement pattern dynamic hypergraph clustering model are carried out to complete the dynamic division of group categories, including:

[0058] Calculate the overall error through the following formula:

[0059]

[0060] Among them, γ1, γ2, and γ3 are respectively And The weight coefficients of;

[0061] Relying on the N individual movement pattern feature matrices X under T time slices, train the dynamic hypergraph depth clustering model to minimize the cumulative error Until The value no longer changes and the model converges. At this time, the training process of the model is completed;

[0062] Input the N individual movement pattern feature matrices X under T + 1 time slices into the trained model T+1 , and determine the individual associated groups found in the test dataset.

[0063] According to the group division method based on individual movement pattern dynamic hypergraph clustering provided by the present invention, after completing the training and testing of the model, it also includes an example verification of the individual associated group category division, specifically including:

[0064] Input the M individual movement pattern feature matrices X' under T time slices T, using the method established in the present invention, the individual associated groups in the verification dataset are discovered. The proposed method is applied to the public transportation passenger card - swiping dataset, and through multiple groups of comparison and ablation experiments, the performance of the proposed method and the division of its internal modules is verified.

[0065] The present invention also provides a group division system based on dynamic hypergraph clustering of individual movement patterns, including:

[0066] A data acquisition module, configured to acquire historical travel record data of N individuals;

[0067] An individual movement pattern feature matrix extraction module, configured to extract D - dimensional features from the historical travel records to quantitatively express the movement patterns of each individual;

[0068] An individual movement pattern hypergraph set construction module, configured to construct a hypergraph set for the current time period according to the current individual movement pattern feature matrix;

[0069] An individual movement pattern deep expression learning module, configured to rely on the Transformer "encoding - decoding" structure to learn the deep expressions of individual movement patterns;

[0070] A neighborhood group movement pattern deep expression learning module, configured to rely on the hypergraph convolutional network model to learn the deep expressions of neighborhood group movement patterns in the local space;

[0071] A dynamic hypergraph depth clustering module, configured to use the attention mechanism to fuse the above - mentioned deep expression forms of individual and neighborhood group movement patterns, and use the bidirectional hypergraph convolutional network to capture the change value of the hypergraph topological structure under the cross - action of forward - time propagation and backward - time propagation, and update its dynamic hypergraph structure;

[0072] A training and testing module, configured to input the historical travel record set of N individuals into the movement pattern dynamic hypergraph clustering model for training and testing of individual associated group category division.

[0073] A verification module, configured to input the historical travel record set of M individuals into the movement pattern dynamic hypergraph clustering model for instance verification of individual associated group category division.

[0074] In the method and system for group division based on dynamic hypergraph clustering of individual movement patterns provided by the present invention, a movement pattern matrix and a movement pattern hypergraph can be constructed based on the travel record set of an individual. Among them, the constructed movement pattern matrix can divide individuals into different categories from the perspective of time and space that "connect two points unidirectionally", and the movement pattern hypergraph can consider the implicit correlation relationship between individuals from the "many-to-many" perspective on the basis of the classification criteria of time and space, and represent the groups with inherent correlation relationships through the neighborhood groups in the hypergraph. In the solution of this application, by accurately identifying the "many-to-many" correlation relationship in the individual travel pattern, the accurate division of group categories is realized. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0076] Figure 1 It is one of the flow diagrams of the method for group division based on dynamic hypergraph clustering of individual movement patterns provided by an embodiment of the present invention;

[0077] Figure 2 It is the second of the flow diagrams of the method for group division based on dynamic hypergraph clustering of individual movement patterns provided by an embodiment of the present invention;

[0078] Figure 3 It is the structural diagram of the system for group division based on dynamic hypergraph clustering of individual movement patterns provided by an embodiment of the present invention; DETAILED DESCRIPTION OF THE EMBODIMENTS

[0079] To make the objectives, technical solutions, and advantages of the present invention clearer, the following will clearly and completely describe the technical solutions in the present invention in conjunction with the drawings in the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Based on the embodiments in the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the protection scope of the present invention.

[0080] Figure 1 It is one of the flow diagrams of the method and system for group division based on dynamic hypergraph clustering of individual movement patterns provided by an embodiment of the present invention.

[0081] Figure 2 It is the second of the flow diagrams of the method and system for group division based on dynamic hypergraph clustering of individual movement patterns provided by an embodiment of the present invention.

[0082] As Figure 1 and Figure 2 shown, this embodiment provides a method for grouping a population based on dynamic hypergraph clustering of individual movement patterns, including:

[0083] Step 101: Obtain the historical travel record set of N individuals, cut the set in the time dimension on a daily basis to obtain the historical travel record sets of N individuals under T consecutive time slices, extract D-dimensional travel features for each individual, and construct a movement pattern feature matrix of N individuals;

[0084] Step 102: For any time slice, regard each individual as a node to obtain N nodes; regard the neighborhood group of an individual as a hyperedge of the hypergraph to obtain N hyperedge sets, and construct an individual movement pattern hypergraph under each time slice;

[0085] Step 103: Construct a dynamic hypergraph clustering model based on movement patterns, that is, relying on the "encoding-decoding" structure of Transformer to learn the deep expression of individual movement patterns; relying on the hypergraph convolutional network model to learn the deep expression of the movement patterns of neighborhood groups in the local space; using the attention mechanism to fuse the above deep expressions of individual and neighborhood group movement patterns, and using the bidirectional hypergraph convolutional network to capture the change value of the hypergraph topology under the cross-action of forward time propagation and backward time propagation, dynamically update the hypergraph structure, and complete the training and testing of the model;

[0086] Step 104: Input the historical travel record set of M individuals into the movement pattern dynamic hypergraph clustering model for instance verification of individual associated group category division.

[0087] The individuals herein can be people. Further, the travel record set may include the travel time and travel location of several people traveling. Further, it may include different means of transportation taken, as well as the time and location of boarding the means of transportation and the time and location of leaving the means of transportation, etc.

[0088] In practical applications, the above travel record set can be obtained from the municipal transportation system, for example, from the database of Beijing public transportation.

[0089] In practical applications, a matrix with N rows and D columns can be constructed, X = [x1, x2, …, x N ∈ R N×D . Among them, x i(i = 1, 2, …, N) represents the i-th individual. The D column refers to the features in the set of travel records corresponding to the individual. The features in the set of travel records can include degree, degree centrality, betweenness centrality, closeness centrality, eigenvector centrality, and clustering coefficient. Among them, the degree is the number of edges directly connected to a certain node (individual), which is a basic indicator for measuring the number of direct connections of a node in the network and reflects the basic connectivity of the node. The betweenness centrality measures the number of times a node appears in all the shortest paths, reflects the ability of the node in controlling the information flow and resource flow, and is an important indicator for evaluating the importance of the node. The closeness centrality measures the average shortest path length from a node to all other nodes in the network, reflects the reachability of the node in the network, and is an indicator for evaluating the superiority of the node's position in the network. The eigenvector centrality not only considers the number of neighbors directly connected to the node but also the importance of the neighbor nodes, comprehensively reflects the influence of the node in the network, and is suitable for evaluating the long-term importance of the node in the network. The clustering coefficient measures the tightness of the connections formed among the neighbor nodes of a node, reflects the local connectivity of the network, and is of great significance for understanding the community structure of the network and the social environment of the node.

[0090] A hypergraph is a generalized graph structure. Compared with a traditional graph structure where an edge can only connect two vertices, in a hypergraph structure, an edge can connect any number of vertices. The advantage of the hypergraph structure is that it can distinguish between hierarchical and non-hierarchical information structures, and any complex information structure can be established from these two types of information.

[0091] In practical applications, the nodes and edge weights of the hypergraph can be determined based on a selected number of individuals, and the individual movement pattern hypergraph G m ={V, E, W} can be constructed. Among them, V, E, and W represent the vertex set, hyperedge set, and edge weight set of the hypergraph respectively. Its incidence matrix The internal element h(v, e) characterizes the membership relationship between the node v and the hyperedge e, and v ∈ e means the node is on the hyperedge. D v and D e are the degree matrices of v and e respectively, which describe the importance of any node or hyperedge in the hypergraph. For the node v ∈ V in the hypergraph, its degree d(v) is the sum of the weights of all hyperedges connected to the vertex v, and the set of d(v) constitutes the node degree matrix D v =[d(v)]. For the hyperedge e ∈ E in the hypergraph, its degree d(e) is the sum of the number of nodes connected by the edge e in the hypergraph. The set of d(e) constitutes the hyperedge degree matrix D e =[d(e)].

[0092] When multiple individuals jointly complete a social activity, their movement patterns show similarity. There may be mutual influence among multiple individuals with similar movement patterns. The more similar the movement patterns are, the higher the degree of mutual influence among individuals. Compared with the traditional graph structure that only describes the correlation between paired nodes, the hypergraph aggregates the interaction correlation information of neighboring multiple nodes, which is more conducive to discovering potential modules in the graph and then modeling the correlation characteristics among multiple individuals. Based on this, the concept of hypergraph is introduced in the solution of this embodiment. By constructing a hypergraph of individual movement patterns, the complex high-order correlation relationship of "many-to-many" between individual movement patterns can be modeled.

[0093] In practical applications, the deep expression of the movement pattern matrix can be carried out based on the learning of transformer time series features. Specifically, a transformer model can be introduced, and through the working principle of the multi-head attention mechanism of the built-in encoder of the transformer model, the deep expression of the individual movement pattern in the convolutional layer of the movement pattern matrix can be autonomously learned.

[0094] A neighborhood is a topological structure on a set. A neighborhood is a special curve. If a certain point is taken as the center point, any open interval is called the neighborhood of that point.

[0095] In this embodiment, operations of a hypergraph convolutional network (HGCN) can be performed to capture the local spatial topological changes of the neighborhood group. Furthermore, for each convolutional layer of the hypergraph convolutional network, the deep expression of the neighborhood group, that is, the second expression, can be determined.

[0096] Clustering is a process of dividing a set of physical or abstract objects into similar object classes. A cluster is a set of data objects. These objects are similar to the objects in the same cluster and different from the objects in other clusters.

[0097] In the method for dividing group categories provided in this embodiment, a movement pattern matrix and a movement pattern hypergraph can be constructed based on the travel record set of individuals. The constructed movement pattern matrix can divide individuals into different categories from the perspective of time plus space of "one-side connecting two points". Further, the movement pattern hypergraph can, on the basis of the classification criteria of time plus space, consider the implicit correlation relationship among individuals from the "many-to-many" perspective, and characterize the groups with internal correlation relationships through the neighborhood groups in the hypergraph. It is precisely by considering the "many-to-many" correlation relationship in the individual travel patterns that the accurate division of group categories is realized in the solution of this application.

[0098] In an exemplary embodiment, extracting D-dimensional travel features for each individual and constructing a movement pattern feature matrix of N individuals includes:

[0099] At any time slice τ, extract D-dimensional travel features for each individual, and construct a mobile pattern feature matrix X of N individuals. τ ∈R N×D , Obtain the individual mobile pattern feature matrix X = {X1, …, X τ , …, X T} under T consecutive time slices, where x τ,i (1 ≤ i ≤ N) is the mobile pattern feature expression of the i-th individual at the τ-th time slice.

[0100] In practical applications, the obtained travel record set must be a travel record within a long time range. In the process of actual group category division, this large-scale travel record set can be sliced by time to obtain several time slices. In implementation, the time slice can be in days. For example, the obtained travel record set can be a complete set of a certain year or a certain month. For example, obtain all the travel records of these individuals in August and divide them by day, such as the travel record set on August 1, the travel record set on August 2, the travel record set on August 3, and so on. By cutting the travel record set in the time dimension, a relatively long time span can be sliced into small time spans. In this embodiment, the group category division in the time dimension is realized in this way.

[0101] After slicing the travel record set of individuals in the time dimension, a historical travel record set of individuals under several consecutive time slices can be obtained. Further, for each time slice, D-dimensional travel features can be extracted for each individual, and a mobile pattern feature matrix X of N individuals can be constructed. τ ∈R N×D , Thus, a sequence of individual mobile pattern feature matrices X = {X1, …, X τ , …, X T} under T consecutive time slices is generated. Among them, x τ,i (1 ≤ i ≤ N) is the mobile pattern feature expression of the i-th individual at the τ-th time slice. Model the association network between individuals based on graph theory. Specifically, for any time slice τ, regard each individual as a node on the graph; connect the connectivity relationship between nodes with edges; use the inverse function of the Euclidean distance between the attributes of two nodes as the corresponding edge weight.

[0102] Further, for any time slice τ, according to the principle that "the smaller the Euclidean distance, the more similar the mobile patterns", circle k individuals with the most similar mobile patterns to form the i-th neighborhood group at the τ-th time slice. To express the complex "many-to-many" association relationship of k individuals at a specific time slice. The obtained neighborhood groups are analogized to hyperedges, and a set of N hyperedges is obtained, forming the initial partitioning result of the individual association group.

[0103] In an exemplary embodiment, the constructing of the individual movement pattern hypergraph for each time slice includes:

[0104] For any time slice τ, regarding the N individuals as the nodes of the hypergraph, obtaining N nodes; using the Euclidean distance to measure the similarity value of the movement patterns of any two individuals; connecting two neighboring nodes, and taking the inverse function of the Euclidean distance as the corresponding edge weight; determining the k individuals with the most similar movement patterns according to the principle that "the smaller the Euclidean distance, the more similar the movement patterns", forming the i-th neighborhood group at the τ-th time slice

[0105] Regarding the neighborhood group as a hyperedge, obtaining a set of N hyperedges, forming the initial partitioning result of the individual association group, and obtaining the individual movement pattern hypergraph G at the τ-th time slice τ =(V τ ,E τ ,X τ ,W τ ,H τ ), where V τ is the node set expressing all individuals, E τ is the hyperedge set expressing the movement patterns of the neighborhood groups, X τ is the node attribute set expressing the individual movement patterns, W τ is the hyperedge weight set expressing the importance degree of the neighborhood groups, H τ is the association matrix expressing the membership relationship between individuals and groups, H τ =[h(v,e)] τ ∈R N×N , h((v,e) represents the membership relationship between node v and hyperedge e;

[0106] The membership relationship between the node v and the hyperedge e conforms to the following formula:

[0107]

[0108] In formula (1), v∈e means the node is on the hyperedge, D τ,v =[d(v)] is the degree matrix of node v at time slice τ, used to describe the importance degree of node v in the hypergraph G τ ; d(v) is the sum of the weights of all hyperedges connected to vertex v, calculated by formula (2):

[0109] d(v)=∑ e∈Eω(e)h(v,e) (2)

[0110] D τ,e = [d(e)] is the degree matrix of the hyperedge e under the time slice τ, which describes the importance of the hyperedge e in the hypergraph G under the time slice τ. d(e) is the sum of the number of all nodes connected by the hyperedge e, and is calculated by the formula (3): τ in which, d(e) is the sum of the number of all nodes connected by the hyperedge e, and is calculated by the formula (3):

[0111] d(e) = ∑ v∈V h(v,e) (3)

[0112] In an exemplary embodiment, the construction of the dynamic hypergraph clustering model based on the movement pattern includes:

[0113] Using the Transformer model to learn the deep representation of the individual movement pattern;

[0114] Using the hypergraph convolutional network model to learn the deep expression of the neighborhood group movement pattern in the local space;

[0115] Fusing the individual and neighborhood group movement patterns for dynamic hypergraph deep clustering;

[0116] Performing the training and testing processes of the dynamic hypergraph clustering model for the movement pattern.

[0117] In an exemplary embodiment, the use of the Transformer model to learn the deep representation of the individual movement pattern includes:

[0118] Through the working principle of the multi-head attention mechanism of the built-in encoder of the Transformer model, autonomously learn the deep expression C of the individual movement pattern in the individual movement pattern feature matrix X in the l-th (1 ≤ l ≤ L) convolutional layer (l) , C (l) is calculated by the formula (4):

[0119]

[0120] where, σ ReLU is the linear activation function. When l = 1, C (l-1) = C 0 is the individual movement pattern feature matrix X; LN is the convolutional layer regularization operation; dropout is the operation of randomly pruning the fully connected layer; FeedForward is the fully connected network, Res (l) is the residual unit of the decoder of the Transformer model; U1 and U2 are the weight parameters to be trained by the model, b1 and b2 are the bias values; MultiHead is the multi-head attention mechanism;

[0121] Through the built-in decoder of the Transformer model, after passing through L convolutional layers, the reconstructed Complete the decoding operation;

[0122] Minimize the average reconstruction error To obtain a deep expression form of the individual movement pattern that is closest to the real situation. In the formula, ||·|| Fro Is the F-norm, where the average reconstruction error is generated during the encoding and decoding processes, and the average reconstruction error The expression is as shown in formula (5):

[0123]

[0124] For any time slice τ, according to the distribution of the feature matrix X τ And the true group category to which each individual belongs within it, calculate the true probability distribution of individual i belonging to category j, and then obtain the probability distribution curve GT of the individual movement pattern category in the real scenario under time slice τ;

[0125] According to the distribution of the deep expression C of the individual movement pattern in the L-th convolutional layer (L) Obtain the probability distribution curve TR of the individual movement pattern category predicted by the Transformer under time slice τ;

[0126] Determine the cumulative error of the probability distribution curve GT and the probability distribution curve TR over all time slices Determine the group category division error after the Transformer temporal feature learning, where Is calculated by formula (6),

[0127]

[0128] Among them, GT τij And TR τij Respectively refer to the probability that individual i belongs to category j in curve GT and curve TR under time slice τ, and KL is the "Kullback-Leibler divergence" distribution curve.

[0129] In an exemplary embodiment, the use of the hypergraph convolutional network model to learn the deep expression of the neighborhood group movement pattern in the local space includes:

[0130] Based on the hypergraph G τ And the initial division result of the individual associated groups, perform hypergraph convolutional network operations to capture the topological changes in the local space of the neighborhood group. For the l-th convolutional layer, the deep expression Z of the neighborhood group movement pattern (l) Is calculated by formula (7):

[0131]

[0132] Among them, Θ is the convolution hyperparameter; when l = 1, Z (l-1) = Z 0 is the feature matrix X; H τ Transponse represents the transposed matrix of H τ .

[0133] In an exemplary embodiment, fusing the individual and neighborhood group movement patterns for dynamic hypergraph depth clustering includes:

[0134] Concatenating the deep expression C(l) of the individual movement pattern and the deep expression Z(l) of the neighborhood group movement pattern through formula (8) to obtain the concatenation result CZ(l) of C(l) and Z(l) under the l-th convolutional layer, where CZ (l) n is the n-th vector sequence of CZ (l) .

[0135] Introduce an attention mechanism to obtain a new expression of the neighborhood group movement pattern under the l-th convolutional layer through formula (9)

[0136]

[0137]

[0138] where q is the query vector, α n is the attention weight of the n-th element, s(CZ(l) n , q) is the attention scoring function, used to measure the similarity between CZ(l) n and the query vector q, and is quantitatively expressed using the dot product model;

[0139] Based on the distribution of, capture the change ΔG τ in the hypergraph topology under the cross-action of forward time propagation and backward time propagation, that is, the change in the hypergraph topology that belongs to the first or second hypergraph but not both at the same time. The calculation method is shown in formula (10):

[0140] ΔG τ = difference(G 1→τ , G T→τ ) = (G 1→τ ∪ G T→τ ) - (G 1→τ ∩

[0141] G T→τ ) (10)

[0142] Based on G τ and ΔG τ , dynamically update the corresponding dynamic hypergraph G τ , which is used as the input of the hypergraph convolutional network in the (l + 1)-th convolutional layer, and perform attention fusion with the corresponding output result C of the Transformer in the (l + 1)-th layer (l+1) , then perform hypergraph convolutional network operations to obtain the probability distribution results of the population categories under the (l + 1)-th to L-th convolutional layers

[0143]

[0144] where W′ τ , H′ τ , D′ τ,v , D′ τ,e are respectively the hyperedge weight set, incidence matrix, node degree matrix, and hyperedge degree matrix of the dynamic hypergraph ;

[0145] For any time slice τ, according to the distribution of the deep expression Z (L) of the individual movement patterns in the l-th convolutional layer, obtain the probability distribution curve HC of the individual movement pattern categories obtained by hypergraph clustering under the time slice τ, and measure the cumulative error between the probability distribution curve HC and the curve GT in all time slices to measure the population category division error of the dynamic hypergraph depth clustering that fuses individual and neighborhood population movement patterns , which is calculated by Equation (12):

[0146]

[0147] where HC τij is the probability that individual i belongs to category j in the curve HC under the time slice τ

[0148] In an exemplary embodiment, the process of training and testing the movement pattern dynamic hypergraph clustering model includes:

[0149] Calculate the overall error through the following formula:

[0150]

[0151] where γ1, γ2, and γ3 are respectively and weight coefficients;

[0152] Relying on the N individual movement pattern feature matrices X under T time slices, train the movement pattern dynamic hypergraph clustering model to minimize the cumulative error until The value no longer changes, and the model converges. At this time, the training process of the model is completed;

[0153] Input the N-bit individual movement pattern feature matrix X under T+1 time slices into the trained model T+1 to determine the individual association groups in the test dataset.

[0154] In an exemplary embodiment, after the training and testing of the model are completed, it further includes instance verification for classifying individual association groups, specifically including:

[0155] Input the M-bit individual movement pattern feature matrix X′ under T time slices T , and use the method established in the present invention to discover the individual association groups in the verification dataset. The proposed method is applied to the public transportation passenger card swiping dataset, and through multiple groups of comparison and ablation experiments, the performance of the proposed method and the division of its internal modules is verified.

[0156] The following uses a specific embodiment to illustrate the group category division method provided by the solution of the present application.

[0157] (1) Working environment

[0158] Use the PyTorch framework to write the model code. All experimental codes are compiled and executed in the Linux software environment, and run on a hardware workstation equipped with an "Intel(R) Xeon(R) CPU E5-2620v4 @ 2.10GHz" CPU and an "NVIDIA P100 16G Tesla" GPU graphics card.

[0159] (2) Introduction to the experimental dataset

[0160] The experimental dataset of the present invention is the public transportation passenger card swiping dataset in a certain place. The transaction period is August 2019, and it is 790,000 travel records generated by 13,962 individuals. The present invention has previously obtained the group categories to which 2,792 (20%) individuals belong. Among them, the first category includes 848 individuals, whose travel time and space are very regular, mainly composed of commuting office workers; the second category includes 690 individuals, whose travel time is regular, but the travel space is irregular, mainly composed of office workers with unfixed work locations; the third category includes 724 individuals, whose travel time is irregular, but the travel space is regular, mainly composed of office workers with unfixed working hours; the fourth category includes 530 individuals, whose travel time and space are both irregular, often composed of deliverymen, domestic workers, repairmen, etc. The travel characteristics of these groups constitute a label matrix to verify the performance of the group category division results.

[0161] (3) Selection of the benchmark model

[0162] In the comparison and ablation experiment session, seven baseline models were selected to compare their performance with the proposed model on the given dataset.

[0163] k-means clustering model: A type of unsupervised learning method that discovers the hidden structure in data by dividing the observation points in the dataset into different groups or clusters.

[0164] AE (Autoencoder): The built-in encoder maps the input data to a low-dimensional latent space, and then clustering tasks are performed in this space.

[0165] IDEC (Improved deep embedded clustering): Improved from Deep Embedded Clustering (DEC), it uses the weights of the pre-trained autoencoder to map the original high-dimensional features to a low-dimensional space and perform clustering tasks in this space while preserving the local structure.

[0166] SDCN (Structured deep clustering network): Constructs a k-nearest neighbor graph based on feature similarity for autoencoder pre-training to capture the aggregation structure relationship between data.

[0167] AGCN (Attention-driven graph clustering network): Extracts multi-scale node attribute and graph topology feature information, performs dynamic fusion, and executes the clustering task of data.

[0168] CaEGCN (Cross-attention Fusion based Enhanced Graph Convolutional Network): Through a deep clustering framework with cross-attention, it fuses the autoencoding module and the graph convolutional module to aggregate heterogeneous data.

[0169] RGCC (Robust Graph Convolutional Clustering): A clustering algorithm that combines a graph convolutional neural network. It does not require prior knowledge of the number of clusters but globally optimizes a continuous objective through a clique convolutional neural network to reduce the dimension and achieve clustering.

[0170] (4) Experimental parameter settings

[0171] The performance study experiment of group division was carried out by using the "AdamW" optimizer to perform iterative training 200 times on a given dataset. The learning rate of all models was set to 10-4. To ensure a fair comparison of all methods, the present invention optimized and set the network configuration parameters recommended according to each baseline model. For K-means, we performed the algorithm 11 times to obtain the best clustering results. The network sizes of AE and IEDC were set to "500-500-2000-10-2000-500-500". The network sizes of the built-in encoder and GCN module of SDCN were both set to "500-500-2000-10". The network sizes of the built-in CAE and GAE modules of CaEGCN were both configured to "500-10-500-500". For the method proposed in the present invention, the weight coefficients γ1, γ2, and γ3 in Formula 11 were set to 1, 0.1, and 0.01 respectively. The built-in AE and HGCN of the proposed method both had a 4-layer network structure, which were set to 500, 500, 2000, and 10 respectively. The total number of iterative training was 200 times.

[0172] (5) Selection of evaluation metrics

[0173] The clustering performance of each model was evaluated through four common metrics, namely accuracy (ACC), normalized mutual information (NMI), average Rand index (ARI), and F1-score (F1). Generally speaking, the larger the evaluation metric, the better the effect.

[0174] (6) Experimental results and analysis

[0175] Table 1 shows the results of the group division comparison experiment of each model on a public transportation passenger card swiping dataset in a certain place. Generally speaking, the proposed method achieved the best clustering performance (the bold data in Table 1). Compared with the second-best baseline model RGCC in terms of group division performance, the proposed method achieved a relatively large increase in each metric, namely: ACC increased by 4.59%, NMI increased by 0.57%, ARI increased by 2.49%, and F1 increased by 4.82%, which confirmed that the present invention has good group division performance. This is attributed to the hypergraph module and feature fusion module of the method proposed in the present invention. The former can fully learn the similarity of the deep movement patterns between individuals and between domain groups; the latter can effectively combine the movement pattern characteristics of individuals themselves and their neighboring groups to better describe the implicit clustering characteristics of target individuals, thereby improving the performance of group division.

[0176] Table 1 Comparison experiment results of each model on a public transportation passenger card swiping dataset in a certain place

[0177]

[0178] Table 2 reflects the improvement of the hypergraph module and feature fusion module of the proposed method on the performance of group division. The data in the second row of the table is the distribution of the group division performance when the proposed method only carries the first module; the data in the third row of the table is the distribution of the group division performance when the proposed method carries the above two modules. It can be seen from the table that adding the hypergraph module can increase the ACC of the proposed model by 1.93%, NMI by 2.08%, ARI by 4.11%, and F1 by 1.90%, which verifies that the hypergraph module can effectively explain the complex many-to-many association characteristics among individuals; after adding the feature fusion module, on the basis of the hypergraph module, the proposed model can further increase the ACC by 2.57%, NMI by 4.81%, ARI by 6.99%, and F1 by 1.48%. This fully shows that the design of the hypergraph module and feature fusion module in the present invention is helpful to improve the accurate division performance of groups.

[0179] Ablation experiment results of the built-in modules of the proposed method on a public transportation passenger card swiping dataset in a certain place

[0180]

[0181] Combined with the above examples, it can be determined that the group category division method provided by this application has at least the following beneficial effects:

[0182] 1) Capture the deep expression characteristics of the movement patterns of individuals and their neighboring groups, use the hypergraph convolutional network for deep association learning and dynamic clustering, realize the accurate division of group categories, and improve the accuracy and robustness of group division;

[0183] 2) Effectively serve the field of public place traffic safety supervision, provide an important theoretical reference for reasonably adjusting the contradiction between traffic supply and demand, effectively allocating urban passenger transport resources, accurately managing the crowded passenger flow in places, effectively ensuring the safe operation of the urban traffic system, and greatly improving the efficiency of public travel, and is expected to provide a more refined group division solution for fields such as urban traffic planning, social media analysis, and disease spread prediction, promoting the development and progress of related fields.

[0184] Next, the group category division system provided by the present invention will be described. The group category division system described below can be mutually corresponding and referred to the group category division method described above.

[0185] Figure 3 It is a schematic structural diagram of a group division system based on dynamic hypergraph deep clustering of movement patterns provided by an embodiment of the present invention.

[0186] As Figure 3 shown, the group division system based on dynamic hypergraph deep clustering of movement patterns provided in this embodiment includes:

[0187] A data acquisition module 301, configured to acquire historical travel record data of N individuals;

[0188] An individual movement pattern feature matrix extraction module 302, configured to extract D-dimensional features from historical travel records to quantitatively express the movement patterns of each individual;

[0189] An individual movement pattern hypergraph set construction module 303, configured to construct a hypergraph set for the current time period according to the current individual movement pattern feature matrix;

[0190] An individual movement pattern deep expression learning module 304, configured to rely on the Transformer "encoding-decoding" structure to learn the deep expressions of individual movement patterns;

[0191] A neighborhood group movement pattern deep expression learning module 305, configured to rely on a hypergraph convolutional network model to learn the deep expressions of neighborhood group movement patterns in the local space;

[0192] A dynamic hypergraph depth clustering module 306, configured to use an attention mechanism to fuse the above-mentioned deep expression forms of individual and neighborhood group movement patterns, and use a bidirectional hypergraph convolutional network to capture the hypergraph topological structure change values under the cross-action of forward time propagation and backward time propagation, and update its dynamic hypergraph structure;

[0193] A training and testing module 307, configured to input the historical travel record set of N individuals into the movement pattern dynamic hypergraph clustering model for training and testing of individual-associated group category division.

[0194] A verification module 308, configured to input the historical travel record set of M individuals into the movement pattern dynamic hypergraph clustering model for instance verification of individual-associated group category division.

[0195] The specific implementation method of the group category division system provided in this embodiment can be implemented with reference to the above-mentioned embodiment, and will not be elaborated here.

[0196] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course also by hardware. Based on such an understanding, the essence of the above technical solution, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disc, etc., including several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods of each embodiment or some parts of the embodiments.

[0197] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for group division based on dynamic hypergraph clustering of individual movement patterns, characterized in that, Including: Obtain the historical travel record set of N individuals, cut the set in the time dimension on a daily basis to obtain the historical travel record sets of N individuals under T consecutive time slices, extract D-dimensional travel features for each individual, and construct a mobile pattern feature matrix of N individuals; For any time slice, regard the individual as a node to obtain N nodes; regard the neighborhood group of the individual as a hyperedge of the hypergraph to obtain N hyperedge sets, and construct an individual mobile pattern hypergraph under each time slice; Construct a dynamic hypergraph clustering model based on the mobile pattern, that is, rely on the Transformer "encoding-decoding" structure to learn the deep expression of the individual mobile pattern; rely on the hypergraph convolutional network model to learn the deep expression of the neighborhood group mobile pattern in the local space; use the attention mechanism to fuse the above deep expressions of the individual and neighborhood group mobile patterns, and use the bidirectional hypergraph convolutional network to capture the change value of the hypergraph topology structure under the cross-action of forward time propagation and backward time propagation, dynamically update the hypergraph structure, and complete the training and testing of the model; Input the historical travel record set of M individuals into the mobile pattern dynamic hypergraph clustering model for instance verification of individual associated group category division; The construction of the individual mobile pattern hypergraph under each time slice includes: For any time slice , regarding the N individuals as nodes of a hypergraph, obtaining nodes; using the Euclidean distance to measure the similarity value of the movement patterns of any two individuals; Connect two adjacent nodes and use the inverse function of the Euclidean distance as the corresponding edge weight; according to the principle that "the smaller the Euclidean distance, the more similar the movement patterns", determine the k individuals with the most similar movement patterns to form the i-th neighborhood group in the i-th time slice ; Taking the neighborhood groups as hyperedges, an N - hyperedge set is obtained, which constitutes the initial partitioning result of the individual association groups, and the individual movement pattern hypergraph at the th time slice is obtained , where is the node set expressing all individuals, is the hyperedge set expressing the movement patterns of neighborhood groups, is the movement pattern feature matrix of N individuals at the th time slice, is the hyperedge weight set expressing the importance of neighborhood groups, is the association matrix expressing the membership relationship between individuals and groups, , characterizes the membership relationship between node and hyperedge .

2. The method for group division based on dynamic hypergraph clustering of individual movement patterns according to claim 1, wherein The extraction of D-dimensional travel features for each individual and the construction of a mobile pattern feature matrix of N individuals include: At any time slice , extract D-dimensional travel features for each individual, and construct a mobile pattern feature matrix for N individuals , , obtain the individual mobile pattern feature matrices under T consecutive time slices , where is the mobile pattern feature expression of the i-th individual under the -th time slice, .

3. The method for group division based on dynamic hypergraph clustering of individual movement patterns according to claim 1, characterized in that The said node and the hyperedge have a membership relationship that conforms to the following formula: （1）； In formula (1), indicates that the node is on the hyperedge, is the node at the time slice The degree matrix below is used to describe the importance of the node in the hypergraph ; is the sum of the weights of all hyperedges connected to the vertex and is calculated by formula (2): （2）； is a hyperedge in the time slice the degree matrix under, to describe the time slice the hyperedge under in the hypergraph the importance degree in, is the sum of the number of all nodes connected by the hyperedge and is calculated by formula (3): （3）。 4. The method for group division based on dynamic hypergraph clustering of individual movement patterns according to claim 1, wherein The construction of a dynamic hypergraph clustering model based on the mobile pattern includes: Utilize the model to learn the deep representation of individual movement patterns; Use the hypergraph convolutional network model to learn the deep expression of the neighborhood group mobile pattern in the local space; Fuse the mobile patterns of the individual and the neighborhood group for dynamic hypergraph deep clustering; Carry out the training and testing process of the mobile pattern dynamic hypergraph clustering model.

5. The method for group division based on dynamic hypergraph clustering of individual movement patterns according to claim 4, wherein The use of a model to learn a deep representation of an individual's movement pattern, including: Through The working principle of the multi-head attention mechanism of the built-in encoder of the model, and autonomously learn the individual movement pattern feature matrix The deep expression of the individual movement pattern in the l-th convolutional layer , Calculated by Equation (4), where : （4）； Among them, is a linear activation function. When l = 1, is the individual movement pattern feature matrix ; is a convolutional layer regularization operation; is an operation of randomly pruning a fully connected layer; is a fully connected network, is the residual unit of the model decoder; and are the weight parameters to be trained in the model respectively, and are bias values; is a multi-head attention mechanism; Through The built-in decoder of the model passes through L convolutional layers to obtain the reconstructed , and completes the decoding operation; Minimize the average reconstruction error , to obtain a depth expression of the individual movement pattern closest to the real situation, where is the Frobenius norm, and the average reconstruction error is generated during the encoding and decoding processes. The expression of the average reconstruction error is shown in Equation (5) as follows: （5）； For any time slice , according to the distribution of the feature matrix and the true group category to which each individual belongs therein, calculate the true probability distribution that individual i belongs to category j, and further obtain the probability distribution curve GT of the individual movement pattern category in the true scenario under the time slice ; According to the deep expression of the individual movement pattern of the L-th convolutional layer to obtain the time slice under the predicted probability distribution curve TR of the individual movement pattern categories; Determine the cumulative error of the probability distribution curve GT and the probability distribution curve TR over all time slices , determine the population category division error after temporal feature learning, where is calculated by Equation (6). （6）； Among them, and respectively refer to the probabilities that individual i belongs to category j in the lower curve GT and curve TR under the time slice The KL is the "Kullback-Leibler divergence" distribution curve.

6. The method for group division based on dynamic hypergraph clustering of individual movement patterns according to claim 5, wherein The use of the hypergraph convolutional network model to learn the deep expression of the neighborhood group mobile pattern in the local space includes: Based on the above hypergraph and the initial partitioning result of the individual associated groups, hypergraph convolutional network operations are performed to capture the local spatial topological changes of the neighborhood groups. For the l-th convolutional layer, the deep expression of the neighborhood group movement pattern is calculated by formula (7): （7）； Among them, is a hypergraph convolutional network, is a convolutional hyperparameter; when l = 1, is the feature matrix X; denotes the transposed matrix of.

7. The method for group division based on dynamic hypergraph clustering of individual movement patterns according to claim 6, wherein Fusing the mobile patterns of the individual and the neighborhood group for dynamic hypergraph deep clustering includes: Deep expression of individual movement pattern With the deep expression of the movement pattern of the neighboring group Concatenated by formula (8) to obtain the result of concatenation under the l-th convolutional layer And The concatenation result , Is The n-th vector sequence of Introduce the attention mechanism, and obtain a new expression for the neighborhood group movement pattern in the l-th convolutional layer through formula (9). : （8）； （9）； Among them, is the query vector, is the attention weight of the nth element, , is the attention scoring function, used to measure the query vector the similarity between them, and is quantitatively expressed using the dot product model; Based on the distribution, capture the hypergraph topological structure changes under the cross-action of forward time propagation and backward time propagation , that is, the hypergraph topological structure changes that belong to the first or second hypergraph but not both at the same time. The calculation method is shown in formula (10): （10）； Based on and , dynamically update the corresponding hypergraph , as the input of the hypergraph convolutional network in the convolutional layer , and perform attention fusion with the output result corresponding to the layer in the layer . After that, perform hypergraph convolutional network operations to obtain the probability distribution result of the population categories under the th convolutional layer ; （11）； Among them, , , , are respectively the hyperedge weight set, incidence matrix, node degree matrix, and hyperedge degree matrix of the dynamic hypergraph ; For any time slice , according to the distribution of the deep expression of the individual movement pattern of the l-th convolutional layer , obtain the individual movement pattern category probability distribution curve HC obtained by hypergraph clustering under the time slice , measure the cumulative error between the probability distribution curve HC and the curve GT in all time slices , and measure the group category division error of the dynamic hypergraph depth clustering that fuses the individual and neighborhood group movement patterns Calculated by Equation (12): （12）； Among them, is the time slice The probability that individual i in the lower curve HC belongs to category j.

8. The method for group division based on dynamic hypergraph clustering of individual movement patterns according to claim 4, wherein The process of training and testing the mobile pattern dynamic hypergraph clustering model includes: Calculate the overall error through the following formula: （13）； Among them, , and are respectively , and 's weight coefficients; Based on the N - bit individual movement pattern feature matrix X under T time slices, train the movement pattern dynamic hyper - graph clustering model to minimize the cumulative error , until the value no longer changes and the model converges. At this time, the training process of the model is completed; Input the N-bit individual movement pattern feature matrix under T+1 time slices into the trained model , and determine the individual association groups found in this test dataset.

9. The method for group division based on dynamic hypergraph clustering of individual movement patterns according to claim 1, wherein After completing the training and testing of the model, it also includes instance verification of individual associated group category division, specifically including: Input the M-bit individual movement pattern feature matrix under T time slices , using the population division method, discover the individual association groups in the validation dataset. The proposed method is applied to the public transportation passenger card swipe dataset, and through multiple groups of comparison and ablation experiments, verify the division performance of the proposed method and its internal modules.

10. A group division system based on dynamic hypergraph clustering of individual movement patterns, applied to the group division method based on dynamic hypergraph clustering of individual movement patterns according to claim 1, characterized in that, Including: A data acquisition module for acquiring the historical travel record data of N individuals; An individual mobile pattern feature matrix extraction module for extracting D-dimensional features from the historical travel records to quantitatively express the mobile pattern of each individual; An individual mobile pattern hypergraph set construction module for constructing a hypergraph set of the current period according to the current individual mobile pattern feature matrix; An individual movement pattern deep expression learning module, which is used to rely on the "encoding-decoding" structure to learn the deep expression of the individual movement pattern; A neighborhood group mobile pattern deep expression learning module for relying on the hypergraph convolutional network model to learn the deep expression of the neighborhood group mobile pattern in the local space; A dynamic hypergraph deep clustering module for using the attention mechanism to fuse the above deep expression forms of the individual and neighborhood group mobile patterns, and using the bidirectional hypergraph convolutional network to capture the change value of the hypergraph topology structure under the cross-action of forward time propagation and backward time propagation, and update its dynamic hypergraph structure; A training and testing module, which is used to input the historical travel record set of N individuals into the mobile pattern dynamic hypergraph clustering model for training and testing of the classification of individual-associated group categories; A verification module, which is used to input the historical travel record set of M individuals into the mobile pattern dynamic hypergraph clustering model for instance verification of the classification of individual-associated group categories.

Images