A virtual track classification method and system based on spatio-temporal semantic fusion

Abstract

The application discloses a kind of virtual track classification method and system of spatiotemporal semantic fusion, based on local dynamic sequencing to virtual track time sequence noise reduction, improve virtual track data quality;Definition of virtual track spatiotemporal semantic expression, virtual track is simplified as time sequence operation and target point, ensure the spatiotemporal continuity of user access process;Propose time sequence operation representation model extraction time sequence operation feature, construct the feature of target point spatial attribute of feature color template extraction, establish POI semantic space extraction target point semantic attribute feature;Using time sequence neural network model, time, space, semantic feature is fused, capture the time sequence dependent relationship of virtual track, realize virtual track accurate classification.The application is helpful to map service provider accurate division user group, establish user interest portrait and develop personalized, intelligent geographic information service.

Description

Technical Field

[0001] This invention belongs to the field of cyberspace information service technology, specifically relating to a virtual trajectory classification method and system based on spatiotemporal semantic fusion. Background Technology

[0002] Public Map Service Platforms (PMSPs), as networked geographic information sharing and service portals, integrate massive amounts of spatial information from multiple sources, providing public geographic information resources to government departments, enterprises, social organizations, and the public. Typical PMSPs include Tianditu, Google Maps, Gaode Maps, and Baidu Maps. The spatial interaction between users and PMSPs creates virtual trajectories, with the trajectory points being the accessed vector map tiles. With the increasing popularity of PMSPs, daily visits in China have exceeded 100 billion, forming a huge user base and diverse needs. However, current map service providers cannot accurately perceive the differences in user preferences and needs, generally providing homogeneous and inefficient services. Therefore, there is an urgent need to establish virtual trajectory classification methods and systems to achieve accurate segmentation of user groups, support the construction of user profiles for map services, and is of great significance for reducing costs, increasing efficiency, and providing personalized, high-quality map services.

[0003] Virtual trajectories differ from typical physical trajectories like GPS trajectories. Due to the pyramidal organization of vector map tiles in PMSP, virtual trajectories also possess a three-dimensional pyramidal spatial structure, with differences in the spatial scale of trajectory points at different levels. In contrast, GPS trajectories are typically two-dimensional trajectories at the same scale. Therefore, although GPS trajectory classification has a broad and relatively mature research system, GPS trajectory classification methods are difficult to apply to virtual trajectory classification. Existing virtual trajectory classification research first extracts target points from the virtual trajectory, then matches the target points with vector map tiles and POIs, assigning spatial and semantic attribute features to the target points. The virtual trajectory is then classified by fusing these two features, with the classification category being the user's domain, such as ocean, river, or forest. This method assumes that users in the same domain have similar access interests; for example, the virtual trajectories of users in the ocean domain are mostly distributed near spatial features such as coastlines and islands. This method has achieved some success. However, this method simplifies the virtual trajectory to target points, losing the user operation information carried by trajectory points between different target points. Furthermore, isolated target points violate the spatiotemporal continuity of the virtual trajectory, ignoring the user's temporal operational behavior habits and limiting the accuracy of virtual trajectory classification. Therefore, there is an urgent need to construct a virtual trajectory classification method and system that integrates multi-dimensional features of time, space, and semantics. Summary of the Invention

[0004] Existing technologies simplify virtual trajectories into isolated target points. While utilizing the spatial and semantic attributes of these target points, they neglect the spatiotemporal continuity of the user's access process and lose the characteristics of the user's temporal operations, resulting in low classification accuracy. To address these issues, this invention proposes a spatiotemporal semantic fusion-based virtual trajectory classification method and system. Based on local dynamic sorting, it reduces temporal noise in virtual trajectory data, improving data quality. It proposes a spatiotemporal semantic representation of virtual trajectories, simplifying them into temporal operations and target points. Models are constructed to extract temporal operation features, spatial attribute features, and semantic attribute features from these components. The fusion of these three features achieves accurate virtual trajectory classification, effectively improving classification accuracy.

[0005] To achieve the above objectives, this invention provides a virtual trajectory classification method based on spatiotemporal semantic fusion, comprising the following steps:

[0006] Step 1: Preprocess the virtual trajectory time series based on local dynamic sorting;

[0007] Step 1.1: Extract vector map tiles accessed by each user from the public map service platform to construct a virtual trajectory;

[0008] Step 1.2: Denoise the virtual trajectory time series based on the local dynamic sorting algorithm;

[0009] Step 2: Express the virtual trajectory in a spatiotemporal semantic manner by defining temporal operations and target points;

[0010] Step 3: Construct a temporal operation representation model and extract temporal operation feature vectors;

[0011] Step 4: Extract the spatial semantic features of the virtual trajectory target points;

[0012] Step 4.1: Construct a ground feature color template to extract the spatial attribute feature vector of the target point;

[0013] Step 4.2: Establish the semantic attribute feature vector of the target point in the POI semantic space;

[0014] Step 5: The temporal operation feature vector extracted in Step 3, the spatial attribute feature vector extracted in Step 4.1, and the semantic attribute feature vector extracted in Step 4.2 are fused to construct a temporal neural network model for virtual trajectory classification based on spatiotemporal semantic feature fusion.

[0015] Furthermore, in step 1.1, each vector map tile represents a trajectory point, the center coordinates of the vector map tile represent the spatial location of the trajectory point, and each trajectory point is described by longitude, latitude, level, and time. The virtual trajectory traj is defined as:

[0016] traj = <p1,p2,p3,…,pi ,…,p n > (1)

[0017] In the formula, p i Let p represent the i-th trajectory point. i =(lon) i ,lat i layer i ,time i ), 1≤i≤n, lon i Indicates longitude, lat i Indicates latitude, layer i Indicates hierarchy, time i This represents time, and n is the number of trajectory points.

[0018] Furthermore, in step 1.2, the virtual trajectory traj is traversed using a window of length k, where the window w is defined as:

[0019] w = <p i ,p i+1 ,…,p i+k-1 > (2)

[0020] In the formula, p i This represents the i-th trajectory point of the virtual trajectory traj.

[0021] Therefore, the cost of the sorted virtual trajectory within this window relative to the original virtual trajectory is defined as:

[0022] cost(i,j)=M(i,j)+L(i,j) (3)

[0023] In the formula, cost(i,j) represents the cost of the trajectory point p. i The sorting cost of the window starting from point P, where M(i,j) represents any trajectory point p within the window. j and starting point p i The movement cost resulting from swapping positions, L(i,j) represents the potential difference cost of the window.

[0024] The calculation method for M(i,j) is as follows:

[0025] M(i,j)=ji,(0 <i<j<k) (4)

[0026] L(i,j) is calculated as follows:

[0027]

[0028] In the formula, i and j are trajectory points p i p j The subscript, k is the window length, l jDenote the trajectory point p in the window j The level of, the potential difference cost L(i,j) represents the trajectory point p j And the starting point p i After swapping positions, the sum of the level differences between adjacent points in the window.

[0029] The process of reordering expects to make the virtual trajectory sequence smooth by exchanging trajectory points within a local range, that is, to minimize the sorting cost. Traverse the virtual trajectory traj with window w, and calculate the iterative cumulative sorting cost as:

[0030] LDS(i) = LDS(i - 1) + min{cost(i,j),(0 < i < j < k)} (6)

[0031] In the formula, min{cost(i,j),(0 < i < j < k)} represents the cost generated by the sorting method with the minimum sorting cost in the window starting from the trajectory point p i , LDS(i) represents the total sorting cost accumulated to the i-th window position. When i = 1, LDS(1) = min(cost(1,j)).

[0032] Moreover, in step 2, the virtual trajectory is simplified to represent the process of target point transfer through temporal operations, including two parts: temporal operations and target points:

[0033] straj = {<op1,tg1>,<op2,tg2>,…,<op i ,tg i ,…,<op r ,tg r >} (7)

[0034] In the formula, straj represents the virtual trajectory after local dynamic sorting, the temporal operation op i Represents the user operation process in the virtual trajectory, and the target point tg i Represents the spatial semantic features preferred by the user in the virtual trajectory, and r represents the number of target points.

[0035] Define the virtual trajectory temporal operation from two perspectives: translation within the same layer and zooming in and out across layers, that is:

[0036]

[0037] In the formula, op i Represents the temporal operation, Pan mi Represents the translation operation within the same layer, Zoom mi Represents the cross-layer operation, and d is the operation length.

[0038] Furthermore, the translation operation within the same layer is described by the translation distance and the translation direction, as defined below:

[0039] Pan mi ={Tiledis,Direction} (9)

[0040] In the formula, Tiledis represents the distance the tile moves during the timing operation, and Direction represents the direction of movement, which is quantized in the four directions of east, south, west, and north.

[0041] Tiledis defines the tile distance as:

[0042] Tiledis = |col p -col q |+|row p -row q | (10)

[0043] In the formula, col p and row p These represent the column and row numbers of the vector map tile p, respectively. q and row q These represent the column number and row number of the vector map tile q, respectively.

[0044] Cross-layer operations do not involve movement distance or direction, only layer changes; therefore, Zoom... mi Defined as:

[0045] Zoom mi ={zoom} (11)

[0046] In the formula, zoom represents the zoom level.

[0047] Pan mi and Zoom mi These are two mutually exclusive basic operations. To unify them, we further define the operation distance Opdis, which represents the number of time-series operations required to move from one target point to the next. The specific calculation formula is as follows:

[0048]

[0049] In the formula, window represents the number of vector map tiles that the browsing window can hold when a user accesses the map.

[0050] Based on the above definitions of temporal operations, the temporal operation sequence Op constructed from the virtual trajectory is represented as follows:

[0051] Op= <op1,op2,…,op i ,…,op r > (13)

[0052] In the formula, op i ={layer i Tiledis i ,Direction i Zoom i Opdis i}, where r represents the number of target points.

[0053] Compared to the basic operations in formula (8), op i Extended attributes have been added.

[0054] The target points in the virtual trajectory are extracted using the HGMM model, and then the target point sequence Tg is constructed:

[0055] Tg = <tg1,tg2,…,tg i ,…,tg r > (14)

[0056] In the formula, tg i Let r represent the i-th target point, which is composed of multiple trajectory points; r represents the number of target points.

[0057] Compared to the basic operations in formula (8), op i Extended attributes have been added.

[0058] Furthermore, the temporal operation representation model in step 3 adopts an Encoder & Decoder structure. The Encoder module transforms the input sequence into a dense vector of fixed dimensions, and the Decoder module decodes the dense vector back into the input sequence. Through repeated learning, the model converges. The dense vector is the temporal operation feature vector. The Encoder module uses a two-layer LSTM to model the temporal features. LSTM is the basic temporal modeling unit. The temporal operation sequence Op is input into the Encoder module, and h is output by the LSTM. end Calculate the distribution parameters, mean μ and variance σ, of the eigenvectors of the time-series operations:

[0059]

[0060] In the formula, W μ With W σ b represents the weight. μ With b σ Indicates deviation, h end This represents the hidden state at the last time step.

[0061] Based on the reparameterization technique, the temporal operation feature vector V is calculated using the mean μ and variance σ. op :

[0062] V op =μ+σε ε~N(0,1) (16)

[0063] In the formula, ε represents the parameter that follows a standard normal distribution.

[0064] The Decoder module and the Encoder module have a symmetrical LSTM structure. The structure of the Decoder module is as follows:

[0065]

[0066] In the formula, This is the initial state of the Decoder module, Op' is the output of the Decoder module, and tanh and s are... i gm is the activation function. This refers to the state of the LSTM in the Decoder module, trained using the Adam optimizer, and its weights W are iteratively updated. z W out and deviation b z b out .

[0067] The loss function consists of two parts: KL divergence and mean squared error, namely:

[0068] loss=KL(N(μ,σ),N(0,1))+MSE(Op,Op') (18)

[0069] KL divergence is used to calculate the distance between the probability distribution N(μ,σ) of the eigenvectors of time-series operations and the normal distribution N(0,1), i.e.:

[0070]

[0071] The MSE loss function requires minimizing the reconstruction error between the input data Op of the Encoder module and the output data Op' of the Decoder module, i.e.:

[0072]

[0073] In the formula, r represents the number of target points, ||Op i -Op' i || 2 means Op i With Op' i The square of the difference.

[0074] Furthermore, in step 4.1, the virtual trajectory target point is overlaid with vector map tiles for analysis. Vector map tiles that spatially overlap with the target point are selected, and a feature color template is constructed. The categories of feature elements on the selected vector map tiles are determined based on RGB values. The proportion of feature element categories in the vector map tiles is extracted to form a spatial attribute feature vector, representing the spatial attribute characteristics of the virtual trajectory. The feature color template describes the RGB color values ​​of different feature types at different levels. By collecting vector map tiles of different feature elements, the RGB color values ​​of the feature elements are identified to form the feature color template. For any vector map tile, the extracted spatial attribute feature vector is:

[0075] V tile =[v1…v e …v τ ] (twenty one)

[0076]

[0077] In the formula, V tile It is a τ-dimensional spatial attribute feature vector, where each dimension represents a type of land cover subclass; This indicates the number of pixels representing that feature subclass within the tile; v e This indicates the pixel percentage of that feature subcategory.

[0078] Furthermore, in step 4.2, the POI context around the virtual trajectory target point is retrieved through a buffer, the average vector of the POI vector set is calculated, and a semantic attribute feature vector is formed to characterize the semantic attribute features of the virtual trajectory target point. The specific operation is as follows: For any POI in the POI database covering POIs nationwide, a buffer with radius R is established with the POI as the center, and the POI context set within the buffer is retrieved. in This indicates the type of POI, and the POIs are sorted in descending order of Euclidean distance. R and u are both hyperparameters, set as needed, and ω... m This is a POI spatial corpus, representing the spatial co-occurrence semantic relationships between different POI types. The Word2vec model is used to train ω. m POI type Transform into POI vector γ u All POI vectors form the POI semantic space; for any target point tg of the virtual trajectory r Establish a radius of R * The buffer is used to construct the target point tg from all POIs within the buffer. r The POI context set, which represents the target point tg rThe semantic attribute information is obtained, and then the POI semantic space is used to transform the POI context set into a POI vector set, and the average vector V of all vectors in the POI vector set is calculated. poi tg, take it as the target point r The semantic attribute feature vector.

[0079] Furthermore, in step 5, the virtual trajectory temporal operation feature vector V is concatenated using vector splicing. op The spatial attribute feature vector V of the virtual trajectory target point tile and semantic attribute feature vector V poi Perform feature fusion:

[0080]

[0081] In the formula, This represents the spatiotemporal semantic fusion features of the trajectory straj at time step t. These represent the temporal operation feature vector, spatial attribute feature vector, and semantic attribute feature vector at time step t, respectively.

[0082] Following the chronological order, the spatiotemporal semantic fusion features of each time step are connected to form the feature sequence of the virtual trajectory straj. r represents the number of target points. This sequence contains the user's temporal, spatial, and semantic information at each time step. The changes in feature values ​​between sequences indicate the evolution of the user's temporal operations and spatial element preferences. A temporal neural network model is constructed using LSTM, with the feature temporal sequence FSeq as input data. Cross-entropy loss is used as the loss function, and the model is trained using the Adam optimizer to perform virtual trajectory classification tasks. During training, two LSTM layers are used to process the FSeq and capture its temporal dependencies. The hidden state and memory unit of the feature temporal sequence FSeq at each time step are calculated to form the hidden state time sequence H =<h(1),…,h(t),…,h(r)> As the output sequence of the LSTM layer, the hidden state h(r) of the last time step is extracted and fed into the fully connected layer to generate the classification prediction result:

[0083] G(straj) = Softmax(W) g h(r)+b g ) (twenty four)

[0084] In the formula, G(straj) represents the classification result of the virtual trajectory straj, Softmax represents the activation function, and W g b represents the weight. g Indicates bias.

[0085] The present invention also provides a spatiotemporal semantic fusion virtual trajectory classification system for implementing the spatiotemporal semantic fusion virtual trajectory classification method described above.

[0086] Furthermore, it includes a processor and a memory, the memory being used to store program instructions, and the processor being used to call the stored instructions in the memory to execute a spatiotemporal semantic fusion virtual trajectory classification method as described above.

[0087] Alternatively, it may include a readable storage medium on which a computer program is stored, which, when executed, implements a spatiotemporal semantic fusion virtual trajectory classification method as described above.

[0088] Compared with the prior art, the present invention has the following advantages:

[0089] 1) This invention uses local dynamic sorting to denoise the virtual trajectory temporally, improving the quality of virtual trajectory data; 2) It defines the spatiotemporal semantic expression of the virtual trajectory, simplifying it into two parts: temporal operations and target points, ensuring the spatiotemporal continuity of the user access process; 3) It proposes a temporal operation representation model to extract temporal operation features, constructs a ground feature color template to extract spatial attribute features of target points, and establishes a POI semantic space to extract semantic attribute features of target points; 4) It uses a temporal neural network model to fuse time, space, and semantic features, capturing the temporal dependencies of the virtual trajectory and achieving accurate classification of the virtual trajectory; 5) This invention is simple and convenient to implement, highly practical, and solves the problems of low practicality and inconvenience in actual application of related technologies. It can improve user experience and has significant market value. Attached Figure Description

[0090] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0091] Figure 1 This is a flowchart of a virtual trajectory classification method based on spatiotemporal semantic fusion according to an embodiment of the present invention.

[0092] Figure 2 This is an example diagram of temporal noise reduction of virtual trajectory using the local dynamic sorting algorithm in an embodiment of the present invention.

[0093] Figure 3 This describes the relationship between timing operations and target points in the virtual trajectory of this invention. Detailed Implementation

[0094] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings and embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0095] Example 1

[0096] like Figure 1 As shown, this embodiment of the invention provides a virtual trajectory classification method based on spatiotemporal semantic fusion, comprising the following steps:

[0097] Step 1: Preprocess the virtual trajectory time series based on local dynamic sorting.

[0098] In PMSP, the spatial interaction process of users forms virtual trajectories, recording the spatiotemporal process of user operations and access to vector map tiles. Both vector map tile loading and access log generation processes have delays, causing uncertainty in the trajectory points within a small temporal range. Specifically, this manifests as frequent cross-layer jumps in the hierarchical attributes of the virtual trajectory, violating map operation rules and creating temporal noise in the virtual trajectory. This invention collects massive amounts of user access logs from PMSP, constructs virtual trajectories based on the temporal relationships of tile access records in the user access logs, and reorders the virtual trajectories based on local dynamic sorting to reduce frequent cross-layer jumps and achieve temporal noise reduction of the virtual trajectory.

[0099] Step 1.1, virtual trajectory construction.

[0100] A virtual trajectory is constructed by extracting vector map tiles accessed by each user from the public map service platform. The center coordinates of the vector map tiles represent the spatial location of the trajectory points, and each trajectory point is described by longitude, latitude, level, and time.

[0101] A virtual trajectory is constructed by extracting vector map tiles accessed by each user from PMSP. Each vector map tile represents a trajectory point, and the center coordinates of the vector map tile represent the spatial location of the trajectory point. Therefore, the virtual trajectory traj is defined as follows:

[0102] traj = <p1,p2,p3,…,p i ,…,p n > (1)

[0103] In the formula, p i Let p represent the i-th trajectory point. i =(lon) i ,lat i layer i,time i ), 1≤i≤n, lon i Let lat represent the longitude of the i-th trajectory point. i The layer represents the latitude of the i-th trajectory point. i Represents the level of the i-th trajectory point, time i Let represent the time of the i-th trajectory point, and n be the number of trajectory points.

[0104] Step 1.2: Denoise the virtual trajectory time series based on the local dynamic sorting algorithm.

[0105] By traversing the virtual trajectory through a moving window, calculating the movement cost and potential difference cost, and minimizing the sorting cost function, dynamic sorting and adjustment of trajectory points can be achieved.

[0106] The temporal noise of virtual trajectories is essentially a local misordering of trajectory points in time. The goal of the local dynamic sorting algorithm is to use a local window to scan the virtual trajectory sequence and reduce the frequent jumps in the hierarchical attributes of the virtual trajectory by dynamically adjusting the order of trajectory points within the window, thereby making the virtual trajectory sequence smoother.

[0107] Because the temporal noise of the virtual trajectory exhibits locality, the sorting process needs to control its scope. Therefore, the virtual trajectory traj is traversed using a window of length k, where the window w is defined as:

[0108] w = <p i ,p i+1 ,…,p i+k-1 > (2)

[0109] In the formula, p i This represents the i-th trajectory point of the virtual trajectory traj, which is also the starting point of the window.

[0110] Therefore, the cost of the sorted virtual trajectory within this window relative to the original virtual trajectory is defined as:

[0111] cost(i,j)=M(i,j)+L(i,j) (3)

[0112] In the formula, cost(i,j) represents the cost of the trajectory point p. i The sorting cost of the window starting from point P, where M(i,j) represents any trajectory point p within the window. j and starting point p i The movement cost resulting from swapping positions, L(i,j) represents the potential difference cost of the window.

[0113] The calculation method for M(i,j) is as follows:

[0114] M(i,j)=ji,(0 <i<j<k) (4)

[0115] In the formula, i and j are the subscripts of the trajectory points p i and p j respectively.

[0116] The movement cost represents the movement cost of replacing the starting trajectory point p i with the long-distance trajectory point p j The greater the distance, the higher the movement cost. Therefore, only the trajectory points with a relatively short distance are more inclined to sort and exchange, so as to ensure the locality of reordering.

[0117] The calculation method of L(i,j) is as follows:

[0118]

[0119] In the formula, i and j are the subscripts of the trajectory points p i and p j respectively, k is the window length, and l j represents the level of the trajectory point p j in the window. The potential difference cost L(i,j) represents the sum of the level differences between adjacent points in the window after the trajectory points p j and the starting point p i are swapped.

[0120] The process of reordering expects to make the virtual trajectory sequence smooth through the exchange of trajectory points within a local range, that is, to minimize the sorting cost. Traversing the virtual trajectory traj with the window w, the cumulative sorting cost calculated iteratively is:

[0121] LDS(i) = LDS(i - 1) + min{cost(i,j), (0 < i < j < k)} (6)

[0122] In the formula, min{cost(i,j), (0 < i < j < k)} represents the cost generated by the sorting method with the minimum sorting cost in the window starting from the trajectory point p i , and LDS(i) represents the total sorting cost accumulated to the i-th window position. When i = 1, LDS(1) = min(cost(1,j)).

[0123] By traversing all the trajectory points in the virtual trajectory traj, the optimal local dynamic sorting result straj can be generated. In the local dynamic sorting method, the movement cost and the potential difference cost are two mutually restrictive conditions. The movement cost controls the sorting range, and the potential difference cost controls the smoothing effect. Usually, the two costs cannot reach the minimum at the same time. Therefore, the local dynamic sorting seeks the balance between the two to achieve the optimal sorting and the noise reduction of the virtual trajectory time series. Figure 2 This is an example of the local dynamic sorting algorithm for noise reduction of the virtual trajectory time series, Figure 2In (a), the solid line represents the hierarchical changes of the original virtual trajectory, while the dashed line represents the hierarchical changes of the virtual trajectory after local dynamic sorting. The solid line shows frequent small fluctuations, while the dashed line is relatively smoother. Figure 2 Figure (b) illustrates the relationship between LDS and window length k. As k increases, LDS tends to plateau, reaching its minimum when k is 7. In this embodiment, k is set to 10.

[0124] Step 2: Spatiotemporal semantic representation of virtual trajectories.

[0125] Users browse maps in PMSP through operations such as panning within the same layer and zooming in / out across layers. Different map levels offer different browsing perspectives: lower-level vector maps have lower tile resolution and larger spatial extent, while higher-level vector maps have higher tile resolution but smaller spatial extent. Existing research indicates that zooming in increases browsing interest, while zooming out decreases it. That is, when users discover a potential browsing target on the map, they tend to zoom in to display high-resolution details, forming target points in the virtual trajectory. These target points represent the set of trajectory points that maximize browsing interest. From the perspective of the user's map browsing process, the virtual trajectory can be simplified as a target point transfer process achieved through temporal operations (zooming in, zooming out, panning, etc.), comprising two parts: temporal operations and target points.

[0126] straj = {<op1,tg1> ,<op2,tg2> ,…, <op i ,tg i >,…, <op r ,tg r >} (7)

[0127] In the formula, straj represents the virtual trajectory after local dynamic sorting, and the timing operation op i This represents the user's operation process in the virtual trajectory, with the target point tg. i The spatial semantic features representing user preferences in the virtual trajectory are denoted by r, where r represents the number of target points. Figure 3 It demonstrates the relationship between temporal operations and target points in a virtual trajectory, with the two alternating to form the virtual trajectory.

[0128] Step 2.1, define the virtual trajectory timing operation.

[0129] The virtual trajectory timing operation is defined from two perspectives: translation within the same layer and zooming in / out across layers.

[0130]

[0131] In the formula, op i Pan represents a timing operation. mi"Zoommi" indicates a translation operation within the same layer, while "Zoommi" indicates a cross-layer operation. The length "d" of different time-series operations is relatively random and generally not the same.

[0132] Same-layer panning refers to a user moving or dragging the map window within the same layer to capture vector map tiles in the new field of view, thereby expanding the browsing area and searching for potential browsing targets, i.e., the next target point. Therefore, same-layer panning can be described by the movement distance and direction, defined as follows:

[0133] Pan mi ={Tiledis,Direction} (9)

[0134] In the formula, Tiledis represents the distance the tile moves during the timing operation, and Direction represents the direction of movement, which is quantized in the four directions of east, south, west, and north.

[0135] Virtual trajectories are three-dimensional, and trajectory points exhibit scale differences across different layers. The Euclidean spatial distance between adjacent trajectory points at lower levels is generally much greater than that between adjacent trajectory points at higher levels. This means that traditional spatial distance measurement methods cannot avoid the scale difference problem across different layers. During user browsing of PMSP, vector map tiles are the basic spatial units. Since vector map tiles are images with fixed pixel dimensions, they are invariant across different layers and only depend on the user's temporal operations. Therefore, this invention utilizes tile distance to calculate the user's movement distance, thus avoiding the scale problem between different layers. Tile distance (Tiledis) is defined as:

[0136] Tiledis = |col p -col q |+|row p -row q | (10)

[0137] In the formula, col p and row p These represent the column and row numbers of the vector map tile p, respectively. q and row q These represent the column number and row number of the vector map tile q, respectively.

[0138] Cross-layer operations refer to zooming in and out of a map. Due to scale differences between layers, the spatial regions of adjacent layers must be contained within each other during cross-layer operations. Therefore, there is no change in distance or direction of movement, only a change in layer hierarchy.

[0139] Zoom mi ={zoom} (11)

[0140] In the formula, zoom represents the zoom level, and in the actual virtual trajectory, zoom... mi It may span multiple levels.

[0141] Pan mi and Zoom mi These are two mutually exclusive basic operations. To unify them, this invention further defines the operation distance Opdis, which represents the number of time-series operations required to move from one target point to the next. The specific calculation formula is as follows:

[0142]

[0143] In the formula, window represents the number of vector map tiles that the browsing window can hold when a user accesses the map. According to existing research, the window size is usually 3.

[0144] When moving across layers (zoom≠0, Tiledis=0), the operation distance Opdis is the number of zoom levels |zoom|. When moving within the same layer (zoom=0, Tiledis≠0), the operation distance Opdis is the ratio of the tile distance to the moving window, Tiledis / window. Therefore, the operation distance can be expressed as the sum of the two.

[0145] Based on the above definitions of temporal operations, this invention constructs a temporal operation sequence Op from a virtual trajectory:

[0146] Op= <op1,op2,…,op i ,…,op r > (13)

[0147] In the formula, op i ={layer i Tiledis i ,Direction i Zoom i Opdis i}, where r represents the number of target points.

[0148] Compared to the basic operations in formula (8), op i Extended attributes have been added.

[0149] Step 2.2, define the virtual trajectory target point.

[0150] Target points are trajectory points with semantic information within a virtual trajectory, sharing similar connotations with stop points on a GPS trajectory. Compared to other trajectory points, target points exhibit stronger user interest. Although target points only possess spatial coordinates, they provide an index for associating spatial attributes such as Points of Interest (POIs) and spatial features, making them crucial for analyzing user interests. This invention utilizes the Hierarchical Gaussian Mixture Model (HGMM) to extract target points from the virtual trajectory, thereby constructing a target point sequence Tg.

[0151] Tg = <tg1,tg2,…,tg i ,…,tg r > (14)

[0152] In the formula, tg i Let r represent the i-th target point, which is composed of multiple trajectory points; r represents the number of target points.

[0153] Step 3: Construct a temporal operation representation model and extract temporal operation feature vectors.

[0154] Based on the variational autoencoder model with Encoder & Decoder structure, a temporal operation representation model is constructed. The temporal operation sequence with irregular length and high data dimension in step 2.1 is transformed into a dense vector with fixed dimension to form a temporal operation feature vector, which is used to represent the temporal distribution pattern of user operation behavior in the virtual trajectory.

[0155] Users in different domains have varying needs for map data at different levels, leading to different temporal operational behaviors. For example, in the field of vehicle navigation, users' temporal operations mainly manifest as high-level panning, while in the field of river (lake) chief systems, spatial element querying and browsing are the main needs, with users tending to repeatedly zoom in and out to query the location of spatial elements such as rivers and lakes. Therefore, mining the temporal operational characteristics of users in different domains is key to achieving accurate classification of domain users. A temporal operation Op represents the process of moving from one target point to another, characterized by irregular temporal length and high data dimensionality, making temporal operation modeling difficult. This invention constructs a temporal operation representation model, using deep learning to transform irregular and high-dimensional temporal operations into dense vectors of fixed dimensions, learning the potential distribution patterns of user operational behaviors, thereby supporting the domain user classification task.

[0156] The temporal operation representation model adopts an Encoder & Decoder structure, which is an unsupervised deep learning architecture suitable for unlabeled data and widely used for dimensionality reduction and feature extraction. In the model design process, the temporal relationships of temporal operation attributes such as layers, zoom levels, movement orientation, tile distance, and operation distance are crucial for feature extraction. Therefore, the basic unit in the Encoder & Decoder structure is chosen to be the temporal model LSTM (Long Short Term Memory).

[0157] The Encoder&Decoder structure is a symmetric deep learning architecture. The Encoder module transforms the input sequence into a dense vector of fixed dimensions, and the Decoder module decodes this dense vector back into the input sequence. Model convergence is achieved through repeated learning. The dense vector represents the temporal operation feature vector. In this invention, the Encoder module utilizes a two-layer LSTM to model temporal features. LSTM is the basic temporal modeling unit. The temporal operation sequence Op is input into the Encoder module, and h is output by the LSTM. end Calculate the distribution parameters, mean μ and variance σ, of the eigenvectors of the time-series operations:

[0158]

[0159] In the formula, W μ With W σ b represents the weight. μ With b σ Indicates deviation, h end This represents the hidden state at the last time step.

[0160] Based on the reparameterization technique, the time-series operation feature vector V op It can be calculated using the mean μ and variance σ as follows:

[0161] V op =μ+σε ε~N(0,1) (16)

[0162] In the formula, ε represents the parameter that follows a standard normal distribution.

[0163] The Decoder module and the Encoder module have a symmetrical LSTM structure. The structure of the Decoder module is as follows:

[0164]

[0165] In the formula, This represents the initial state of the Decoder module, Op' is the output of the Decoder module, and tanh and sigm are the activation functions. This refers to the state of the LSTM in the Decoder module, trained using the Adam optimizer, and its weights W are iteratively updated. z W out and deviation b z b out .

[0166] The temporal operational representation model is actually an application of the variational autoencoder model in the field of temporal modeling. The loss function consists of two parts: KL divergence and mean squared error (MSE).

[0167] loss=KL(N(μ,σ),N(0,1))+MSE(Op,Op') (18)

[0168] KL divergence is generally used to measure the distance between two probability distribution functions. Specifically, it is used to calculate the distance between the probability distribution N(μ,σ) of the eigenvectors of a time-series operation and the normal distribution N(0,1), i.e.:

[0169]

[0170] The MSE loss function requires minimizing the reconstruction error between the input data Op of the Encoder module and the output data Op' of the Decoder module, i.e.:

[0171]

[0172] In the formula, r represents the number of target points, ||Op i -Op' i || 2 means Op i With Op' i The square of the difference.

[0173] Step 4: Extract the spatial semantic features of the virtual trajectory target points.

[0174] Vector map tiles, derived from vector maps, contain line and polygon features and are frequently accessed and widely used. Matching virtual trajectory target points with vector map tiles and Points of Interest (POIs) extracts spatial features, characterizing the spatial and semantic attributes of the target points.

[0175] Step 4.1: Extract the spatial attribute feature vector of the virtual trajectory target point.

[0176] The virtual trajectory target point is overlaid with vector map tiles for analysis. Vector map tiles that spatially overlap with the target point are selected. A land feature color template is constructed to extract the proportion of land feature categories in the vector map tiles, forming a spatial attribute feature vector to represent the spatial attribute characteristics of the virtual trajectory.

[0177] The present invention crawls vector map tiles that have a spatial inclusion relationship with the virtual trajectory target points and extracts the spatial attribute features of the target points. The vector map tiles contain line and surface feature information, and the spatial attributes of different ground objects are identified by colors. Compared with remote sensing tiles, it is less difficult and more accurate to extract spatial attributes through the colors of vector map tiles.

[0178] The present invention extracts the spatial attributes of vector map tiles by constructing a ground object color template. The ground object color template describes the RGB color values of different ground object types at different levels. By collecting vector map tiles of different ground object elements and identifying the RGB color values of the ground object elements, this template is formed. Generally speaking, the ground object color templates of different public map service platforms are not the same. The present invention takes Tianditu as the research object. According to the ground object classification standard of Tianditu, the template includes 5 major ground object categories, namely roads, water, green spaces, land, and residential areas, as well as 10 minor ground object categories. Considering the problem of hierarchical display of minor ground object categories during the map generalization process, for example, the functional areas only appear in the [level codes 14, 15, 16, 17, 18], therefore, the extraction ranges of different minor ground object categories are also divided by level. The template is identified by RGB. Considering the slight differences in the colors of vector map tiles, an error range of ±2 is adopted. The specific template is shown in Table 1.

[0179] Table 1 Ground Object Color Template of Vector Map Tiles in Tianditu (2023)

[0180]

[0181] For any vector map tile tile, the extracted spatial attribute feature vector is:

[0182] V tile =[v1 … v e … v 10 (21)

[0183]

[0184] In the formula, V tile [[ID=2,9]]is a 10-dimensional spatial attribute feature vector, and each dimension represents a minor ground object category; represents the number of pixels of this minor ground object category in the tile; v e represents the pixel proportion of this minor ground object category.

[0185] Due to color errors, there are generally a small number of pixels that cannot be accurately identified. Therefore

[0186] Step 4.2, extract the semantic attribute feature vector of the virtual trajectory target point.

[0187] The context of Points of Interest (POIs) around the virtual trajectory target point is retrieved by the buffer, and the average vector of the POI vector set is calculated to form a semantic attribute feature vector, which represents the semantic attribute features of the virtual trajectory target point.

[0188] The semantic attributes of virtual trajectories represent the category attributes of the user's access area corresponding to the trajectory points. This invention utilizes the spatial co-occurrence semantic representation of POIs. First, based on the spatial co-occurrence relationship of POIs, a POI semantic space is constructed using POI data at a national scale to vectorize the semantic attributes of POIs. Then, the POIs matched by trajectory points are represented as feature vectors that encompass semantic attributes.

[0189] For any POI in a POI database covering all POIs nationwide, establish a buffer with radius R centered on that POI, and retrieve the POI context set within the buffer. in This indicates the type of POI, and the POIs are sorted in descending order of Euclidean distance. R and u are both hyperparameters that can be preset as needed. In this embodiment, R is set to 1 km, and u is set to 30. ω m This is a POI spatial corpus, representing the spatial co-occurrence semantic relationships between different POI types, such as the widespread spatial proximity dependency relationship between restaurants and shopping malls.

[0190] To achieve a quantitative representation of co-occurrence semantic relationships in the POI space, the Word2vec model is further used to train ω. m POI type Transform into POI vector γ u For any two POI vectors, the smaller their cosine similarity, the closer the vectors are, and the higher the probability of co-occurrence of the two types of POI spaces. All POI vectors form the POI semantic space.

[0191] For any target point tg on the virtual trajectory r Establish a radius of R * The buffer is used to construct the target point tg from all POIs within the buffer. r The POI context set, which represents the target point tg r The semantic attribute information is obtained, and then the POI semantic space is used to transform the POI context set into a POI vector set, and the average vector V of all vectors in the POI vector set is calculated. poi tg, take it as the target point r The semantic attribute feature vector.

[0192] Step 5: Construct a temporal neural network model to perform virtual trajectory classification by fusing spatiotemporal semantic features.

[0193] By concatenating the temporal operation feature vectors from step 3, the spatial attribute feature vectors from step 4.1, and the semantic attribute feature vectors from step 4.2, a temporal neural network model is constructed to capture the temporal dependencies of virtual trajectories, classify and output the domain categories of virtual trajectories, and realize the domain classification of virtual trajectories.

[0194] The temporal operation sequence Op and the target point sequence Tg of the virtual trajectory strak together constitute the user access process. Users within the same domain share common characteristics in their temporal operations and access to vector map tiles and POIs, while different domains exhibit differences. In this embodiment, agriculture, navigation, environmental protection, rivers, and oceans—domains with a large user base in the top-bottom map—were selected as classification domains.

[0195] In step 3 of this invention, the virtual trajectory temporal operation feature vector V is extracted. op In step 4, the spatial attribute feature vector V of the virtual trajectory target point was extracted. tile and semantic attribute feature vector V poi This invention achieves domain classification of virtual trajectories by fusing three features. Since all three features are vectors, the invention uses vector concatenation for feature fusion.

[0196]

[0197] in, This represents the spatiotemporal semantic fusion features of the trajectory straj at time step t. These represent the temporal operation feature vector, spatial attribute feature vector, and semantic attribute feature vector at time step t, respectively.

[0198] Following the chronological order, the spatiotemporal semantic fusion features of each time step are connected to form the feature sequence of the virtual trajectory straj. r represents the number of target points. This sequence contains the user's temporal, spatial, and semantic information at each time step. The changes in feature values ​​between sequences indicate the evolution of the user's temporal operations and spatial element preferences. Therefore, extracting and learning the transfer relationships of spatiotemporal semantic fusion features in virtual trajectories is the key to capturing user behavior habits and preferences, and is also the basis for achieving accurate classification.

[0199] This invention utilizes a classic LSTM to construct a temporal neural network model, using the feature temporal sequence FSeq as input data to train the model for a virtual trajectory classification task. During training, two LSTM layers are used to process the FSeq and capture its temporal dependencies, calculating the hidden state and memory unit for each time step (target point) of the feature temporal sequence FSeq, forming a hidden state time series H =<h(1),…,h(t),…,h(r)> This is used as the output sequence of the LSTM layer. The hidden state h(r) of the last time step (target point) is extracted and fed into the fully connected layer to generate the classification prediction result:

[0200] G(straj) = Softmax(W) g h(r)+b g ) (twenty four)

[0201] In the formula, G(straj) represents the classification result of the virtual trajectory straj, Softmax represents the activation function, and W g b represents the weight. g Indicates bias.

[0202] The model is trained using cross-entropy loss as the loss function and the Adam optimizer is used.

[0203] Example 2

[0204] Based on the same inventive concept, the present invention also provides a virtual trajectory classification system based on spatiotemporal semantic fusion, including a processor and a memory. The memory is used to store program instructions, and the processor is used to call the program instructions in the memory to execute the virtual trajectory classification method based on spatiotemporal semantic fusion as described above.

[0205] Example 3

[0206] Based on the same inventive concept, the present invention also provides a virtual trajectory classification system with spatiotemporal semantic fusion, including a readable storage medium on which a computer program is stored. When the computer program is executed, it implements the virtual trajectory classification method with spatiotemporal semantic fusion as described above.

[0207] In specific implementation, the method proposed in the technical solution of this invention can be automatically executed by those skilled in the art using computer software technology. System devices for implementing the method, such as computer-readable storage media storing the corresponding computer program of the technical solution of this invention and computer equipment including the computer program running the corresponding computer program, should also be within the protection scope of this invention.

[0208] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to replace them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

Claims

1. A virtual trajectory classification method based on spatiotemporal semantic fusion, characterized in that, Includes the following steps: Step 1: Preprocess the virtual trajectory time series based on local dynamic sorting; Step 1.1: Extract vector map tiles accessed by each user from the public map service platform to construct a virtual trajectory; Step 1.2: Denoise the virtual trajectory time series based on the local dynamic sorting algorithm; With length as Window traversal of virtual trajectory , among which windows Defined as: （2） In the formula, Representing virtual trajectory The A trajectory point; Therefore, the cost of the sorted virtual trajectory within this window relative to the original virtual trajectory is defined as: （3） In the formula, Represented by trajectory points The sorting cost of the window starting from [the specified point]. Represents any trajectory point within the window. and starting point The cost of movement resulting from swapping positions. Represents the potential cost of the window; The calculation method is as follows: （4） The calculation method is as follows: （5） In the formula, For trajectory points , subscript, For window length, Represents the trajectory points in the window The level, potential difference cost Represents trajectory points and starting point The sum of the hierarchical differences between adjacent points within the window after swapping their positions; The reordering process aims to smooth the virtual trajectory sequence by exchanging trajectory points within a local range, i.e., to minimize the sorting cost, using a window... Traversing the virtual trajectory The cumulative sorting cost of the iterations is calculated as follows: （6） In the formula, Represented by trajectory points The cost of the sorting method that minimizes the sorting cost within the window starting from which the sorting occurs. Indicates the cumulative total up to the [number]. The total sorting cost for each window position hour, ; Step 2: Express the virtual trajectory in a spatiotemporal semantic manner by defining temporal operations and target points; The virtual trajectory can be simplified as a process of transferring the target point through time-series operations, including two parts: time-series operations and the target point. （7） In the formula, Represents the virtual trajectory after local dynamic sorting, time-series operations. This represents the user's operation process and target point in the virtual trajectory. Spatial semantic features representing user preferences in virtual trajectories. Indicates the number of target points; The virtual trajectory timing operation is defined from two perspectives: translation within the same layer and zooming in / out across layers. （8） In the formula, Indicates timing operations, This indicates a translation operation within the same layer. Indicates cross-layer operation. For operation length; Furthermore, the translation operation within the same layer is described by the translation distance and the translation direction, as defined below: （9） In the formula, This indicates the distance the tile moves during the timing operation. Indicates the direction of movement, quantified by the four cardinal directions: east, south, west, and north. Tile distance Defined as: （10） In the formula, and These represent vector map tiles. Column and row numbers, and These represent vector map tiles. Column and row numbers; Cross-layer operations do not involve movement distance or direction, only layer changes, therefore... Defined as: （11） In the formula, Indicates the zoom level; and These are two mutually exclusive basic operations. To unify them, the operation distance is further defined. This represents the number of timing operations required to move from one target point to the next, and the specific calculation formula is as follows: （12） In the formula, This indicates the number of vector map tiles that the browsing window can hold when a user accesses the map; Based on the above definitions of temporal operations, a sequence of temporal operations is constructed from the virtual trajectory. It is expressed as follows: （13） In the formula, , Indicates the number of target points; Compared to the basic operations in formula (8), Extended attributes have been added; The target points in the virtual trajectory are extracted using the HGMM model, and then a sequence of target points is constructed. : （14） In the formula, Indicates the first i A target point is composed of multiple trajectory points; Indicates the number of target points; Step 3: Construct a temporal operation representation model and extract temporal operation feature vectors; Step 4: Extract the spatial semantic features of the virtual trajectory target points; Step 4.1: Construct a ground feature color template to extract the spatial attribute feature vector of the target point; Step 4.2: Establish the semantic attribute feature vector of the target point in the POI semantic space; Step 5: The temporal operation feature vector extracted in Step 3, the spatial attribute feature vector extracted in Step 4.1, and the semantic attribute feature vector extracted in Step 4.2 are fused to construct a temporal neural network model for virtual trajectory classification based on spatiotemporal semantic feature fusion.

2. The virtual trajectory classification method based on spatiotemporal semantic fusion as described in claim 1, characterized in that: In step 1.1, each vector map tile represents a trajectory point. The center coordinates of the vector map tile represent the spatial location of the trajectory point. Each trajectory point is described by longitude, latitude, level, and time. (Virtual trajectory) Defined as: （1） In the formula, Indicates the first A trajectory point, , , Indicates longitude. Indicates latitude, Indicates hierarchy, Indicates time, The number of trajectory points.

3. The virtual trajectory classification method based on spatiotemporal semantic fusion as described in claim 1, characterized in that: In step 3, the temporal operation representation model adopts an Encoder & Decoder structure. The Encoder module transforms the input sequence into a dense vector of fixed dimensions, and the Decoder module decodes the dense vector back into the input sequence. Through repeated learning, the model converges. The dense vector is the temporal operation feature vector. The Encoder module uses a two-layer LSTM to model the temporal features. LSTM is the basic temporal modeling unit, which transforms the temporal operation sequence... The input is fed into the Encoder module and output through LSTM. Calculate the mean of the distribution parameters of the eigenvectors of time-series operations. and variance : （15） In the formula, and Indicates weight, and Indicates deviation, This represents the hidden state at the last time step; Based on the reparameterization technique, through the mean... and variance Calculate the eigenvectors of time-series operations : , （16） In the formula, Parameters representing a standard normal distribution; The Decoder module and the Encoder module have a symmetrical LSTM structure. The structure of the Decoder module is as follows: （17） In the formula, This is the initial state of the Decoder module. This is the output of the Decoder module. and It is an activation function. This refers to the state of the LSTM in the Decoder module, trained using the Adam optimizer, with weights iteratively updated. , and deviation ; loss function It includes two parts: KL divergence and mean square error, namely: （18） KL divergence is used to calculate the probability distribution of eigenvectors of time-series operations. Compared with normal distribution The distance between them, that is: （19） The MSE loss function requires input data from the Encoder module. Output data of the Decoder module Minimize the reconstruction error between them, that is: （20） In the formula, Indicates the number of target points. express and The square of the difference.

4. The virtual trajectory classification method based on spatiotemporal semantic fusion as described in claim 1, characterized in that: In step 4.1, the virtual trajectory target point is overlaid with vector map tiles for analysis, vector map tiles that spatially overlap with the target point are selected, a land feature color template is constructed, the category of land feature elements on the selected vector map tiles is determined based on RGB values, the proportion of land feature element categories in the vector map tiles is extracted, and a spatial attribute feature vector is formed to represent the spatial attribute characteristics of the virtual trajectory. The feature color template describes the RGB color values ​​of different feature types at different levels. By collecting vector map tiles of different feature elements, the RGB color values ​​of the feature elements are identified to form the feature color template. For any vector map tile The extracted spatial attribute feature vector is: （21） （22） In the formula, yes A 3D spatial attribute feature vector, where each dimension represents a type of land cover subclass; This indicates the number of pixels representing that feature subclass within the tile; This indicates the pixel percentage of that feature subcategory. .

5. The virtual trajectory classification method based on spatiotemporal semantic fusion as described in claim 1, characterized in that: In step 4.2, the POI context around the virtual trajectory target point is retrieved through the buffer, and the average vector of the POI vector set is calculated to form a semantic attribute feature vector, which represents the semantic attribute features of the virtual trajectory target point. The specific operation is as follows: For any POI in the POI database covering POIs nationwide, a radius of 10 ... The buffer retrieves the set of POI contexts within it. ,in This indicates the type of POI, and the POIs are sorted in descending order of Euclidean distance. and These are all hyperparameters; set them as needed. This is a POI spatial corpus, representing the spatial co-occurrence semantic relationships between different POI types, trained using a Word2vec model. POI type Convert to POI vector All POI vectors form the POI semantic space; for any target point on the virtual trajectory Establish a radius of The buffer will form the target point by including all POIs within it. The POI context set, which represents the target point The semantic attribute information is obtained, and then the POI semantic space is used to transform the POI context set into a POI vector set, and the average vector of all vectors in the POI vector set is calculated. Take it as the target point The semantic attribute feature vector.

6. The virtual trajectory classification method based on spatiotemporal semantic fusion as described in claim 1, characterized in that: In step 5, the virtual trajectory temporal operation feature vector is concatenated. Spatial attribute feature vector of virtual trajectory target point and semantic attribute feature vector Perform feature fusion: （23） In the formula, Representing the trajectory At time step The spatiotemporal semantic fusion features Representing time steps The temporal operation feature vector, spatial attribute feature vector, and semantic attribute feature vector; By connecting the spatiotemporal semantic fusion features of each time step in chronological order, a virtual trajectory is formed. Feature sequences , This sequence represents the number of target points and contains the user's temporal, spatial, and semantic information at each time step. Changes in feature values ​​between sequences indicate the evolution of the user's temporal operations and spatial preference. An LSTM temporal neural network model is constructed to integrate these features. As input data, cross-entropy loss is used as the loss function, and the Adam optimizer is used to train the model for virtual trajectory classification tasks. During training, two LSTM layers are used for processing. It also captures their temporal dependencies and calculates feature temporal sequences. The hidden state and memory unit at each time step form a hidden state time series. As the output sequence of the LSTM layer, extract the hidden state of the last time step. Entering the fully connected layer, classification prediction results are generated: （24） In the formula, Representing virtual trajectory The classification results This represents the activation function. Indicates weight, Indicates bias.

7. A virtual trajectory classification system that integrates spatiotemporal semantics, characterized in that, It includes a processor and a memory, the memory being used to store program instructions, and the processor being used to call the program instructions in the memory to execute a virtual trajectory classification method based on spatiotemporal semantic fusion as described in any one of claims 1-6.

8. A virtual trajectory classification system that integrates spatiotemporal semantics, characterized in that, It includes a readable storage medium on which a computer program is stored, and when the computer program is executed, it implements a virtual trajectory classification method based on spatiotemporal semantic fusion as described in any one of claims 1-6.

Images