Space-time semantic fusion virtual trajectory classification method and system

Abstract

The invention discloses a space-time semantic fusion virtual track classification method and system, which are used for reducing noise of a virtual track time sequence based on local dynamic sorting and improving the quality of virtual track data. Space-time semantic expressions of the virtual tracks are defined, the virtual tracks are simplified into time sequence operation and target points, and the space-time continuity of the user access process is ensured; a time sequence operation characterization model is provided to extract time sequence operation features, a ground feature color template is constructed to extract target point space attribute features, and a POI semantic space is established to extract target point semantic attribute features; and fusing time, space and semantic features by using a time sequence neural network model, capturing a time sequence dependency relationship of the virtual trajectory, and realizing accurate classification of the virtual trajectory. According to the method, a map service provider can accurately divide user groups, establish user interest portraits and develop personalized and intelligent geographic information services.

Description

Technical Field

[0001] The present invention belongs to the technical field of cyberspace information services, and in particular relates to a virtual trajectory classification method and system for spatiotemporal semantic fusion. Background Art

[0002] As a networked geographic information sharing and service portal, the Public Map Service Platform (PMSP) integrates multi-source massive spatial information and provides geographic information public service resources for government departments, enterprises, institutions, social groups, and the public. Typical PMSPs include Tiandi Map, Google Maps, Amap, Baidu Maps, etc. The spatial interaction process between users and PMSPs forms a virtual trajectory, and the trajectory points are the visited vector map tiles. With the increasing popularity of PMSPs, the daily visits to PMSPs in China have exceeded 100 billion times, forming a huge user group and differentiated needs. However, current map service providers are unable to accurately perceive the differences in user preferences and needs, and generally provide homogeneous and inefficient services. Therefore, it is urgent to establish a virtual trajectory classification method and system to achieve accurate division of user groups and support the construction of user portraits for map services, which is of great significance for reducing costs and increasing efficiency and providing personalized high-quality map services.

[0003] Virtual trajectories are different from typical physical trajectories such as GPS trajectories. Due to the pyramid organization structure of vector map tiles in PMSP, virtual trajectories also have a three-dimensional pyramid spatial structure, and the spatial scales of trajectory points at different levels are different, while GPS trajectories are usually two-dimensional trajectories at the same scale. Therefore, although there is an extensive and relatively mature research system for GPS trajectory classification, GPS trajectory classification methods are difficult to apply to virtual trajectory classification. In existing virtual trajectory classification studies, scholars first extract target points in virtual trajectories, then match target points with vector map tiles and POIs, give target points spatial attribute features and semantic attribute features, and classify virtual trajectories by fusing the two features. The classification category is the user's field, such as ocean, river, forest, etc. The premise of this method is that users in the same field have similar access interests. For example, the virtual trajectories of users in the ocean field are mostly distributed near spatial elements such as coasts and islands. The method has achieved certain results. However, this method simplifies virtual trajectories into target points, loses the user operation information carried by trajectory points between different target points, and isolated target points violate the spatiotemporal process continuity of virtual trajectories, ignores the characteristics of user temporal operation behavior habits, and limits the accuracy of virtual trajectory classification. Therefore, it is urgent to build a virtual trajectory classification method and system that integrates multi-dimensional features of time, space, and semantics. Summary of the invention

[0004] The existing technology simplifies virtual trajectories into isolated target points. Although it utilizes the spatial attribute characteristics and semantic attribute characteristics of the target points, it ignores the spatiotemporal continuity of the user access process and loses the user's temporal operation characteristics, resulting in low virtual trajectory classification accuracy. In view of the above problems, the present invention proposes a virtual trajectory classification method and system with spatiotemporal semantic fusion, which reduces the virtual trajectory temporal noise based on local dynamic sorting, improves the quality of virtual trajectory data, proposes the spatiotemporal semantic expression of virtual trajectories, simplifies virtual trajectories into two parts: temporal operation and target points, and constructs models to extract temporal operation characteristics, spatial attribute characteristics, and semantic attribute characteristics respectively. The three are integrated to achieve accurate classification of virtual trajectories, effectively improving the classification accuracy of virtual trajectories.

[0005] In order to achieve the above object, the present invention provides a virtual trajectory classification method of spatiotemporal semantic fusion, comprising the following steps:

[0006] Step 1, preprocessing the virtual trajectory timing based on local dynamic sorting;

[0007] Step 1.1, extract the vector map tiles visited by each user in the public map service platform to construct a virtual track;

[0008] Step 1.2, denoising the virtual trajectory time series based on a local dynamic sorting algorithm;

[0009] Step 2, express the spatiotemporal semantics of the virtual trajectory by defining the timing operations and target points;

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

[0011] Step 4, extracting the spatial semantic features of the virtual trajectory target point;

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

[0013] Step 4.2, establish POI semantic space to extract semantic attribute feature vectors of target points;

[0014] In 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 with fusion of spatiotemporal and semantic features.

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

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

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

[0018] Moreover, in step 1.2, the virtual trajectory traj is traversed with 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 Represents the i-th trajectory point of the virtual trajectory traj.

[0021] Then, the cost of the sorted virtual trajectory within the 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 trajectory point p i is the sorting cost of the window at the starting point, M(i,j) represents any trajectory point p in the window j and starting point p i The movement cost caused by swapping positions, L(i,j) represents the potential difference cost of the window.

[0024] The calculation method of M(i,j) is:

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

[0026] The calculation method of L(i,j) is:

[0027]

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

[0029] The reordering process expects to make the virtual trajectory sequence smooth by swapping trajectory points within a local range, that is, to minimize the sorting cost. Traverse the virtual trajectory traj with window w, and calculate the cumulative sorting cost of each iteration 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 up 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 on the same layer is described by the moving distance and moving direction, which are defined as follows:

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

[0040] Where Tiledis represents the tile distance moved during the timing operation, and Direction represents the moving direction, which is quantified in four directions: east, south, west, and north.

[0041] The tile distance Tiledis is defined as:

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

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

[0044] Cross-layer operations do not involve moving distance and direction, only layer changes, so Zoom mi Defined as:

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

[0046] In the formula, zoom indicates the level of enlargement or reduction.

[0047] Pan mi and Zoom mi These are two mutually exclusive basic operations. In order to unify the two, the operation distance Opdis is further defined to represent the number of sequential operations required from one target point to the next target point. 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 accommodate when the user accesses the map.

[0050] Based on the above definition of timing operations, the timing operation sequence Op constructed from the virtual trace is expressed 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}, r represents the number of target points.

[0053] Compared with the basic operation in formula (8), op i Added extended attributes.

[0054] The HGMM model is used to extract the target points in the virtual trajectory, and then the target point sequence Tg is constructed:

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

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

[0057] Compared with the basic operation in formula (8), op i Added extended attributes.

[0058] Moreover, the temporal operation representation model in step 3 adopts the Encoder&Decoder structure. The Encoder module converts the input sequence into a dense vector of fixed dimension, and then the Decoder module decodes the dense vector into the input sequence. After repeated learning, the model convergence is achieved, wherein the dense vector is the temporal operation feature vector. The Encoder module uses a double-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 the h output by LSTM is used to generate the temporal features. end Calculate the distribution parameter mean μ and variance σ of the time series operation feature vector:

[0059]

[0060] Where W μ With W σ represents the weight, b μ With b σ Denotes the deviation, h end represents the hidden state at the last time step.

[0061] Based on the reparameterization technique, the time series operation feature vector V is calculated by 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 obeys the 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, is the initial state of the Decoder module, Op' is the output of the Decoder module, tanh and s i gm is the activation function, is the state of LSTM in the Decoder module, trained using the Adam optimizer, and iteratively updates the weight W z , W out and deviation b z , b out .

[0067] The loss function loss includes two parts: KL divergence and mean square 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 time series operation feature vector and the normal distribution N(0,1), that is:

[0070]

[0071] The MSE loss function requires that the reconstruction error between the input data Op of the Encoder module and the output data Op' of the Decoder module be minimized, that is:

[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] Moreover, in step 4.1, the virtual trajectory target point and the vector map tile are overlaid and analyzed, the vector map tiles that overlap with the target point in space are screened, a feature color template is constructed, the category of the feature element on the screened vector map tile is determined according to the RGB value, the feature element category ratio of the vector map tile is extracted, and a spatial attribute feature vector is formed to characterize 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 a 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] Where V tile It is a τ-dimensional spatial attribute feature vector, where each dimension represents a small category of ground features; Indicates the number of pixels of this feature class in the tile; v e Indicates the pixel ratio of the small class of the feature.

[0078] Moreover, in step 4.2, the POI context around the virtual trajectory target point is retrieved through the buffer, the average vector of the POI vector set is calculated, and a semantic attribute feature vector is formed to characterize the semantic attribute characteristics of the virtual trajectory target point. The specific operation is as follows: for any POI in the POI database covering POIs across the country, a buffer with a radius of R is established with the POI as the center, and the POI context set in the buffer is retrieved in Indicates the type of POI, and POIs are sorted from large to small according to the Euclidean distance. R and u are both hyperparameters, which can be set as needed. ω m It is the POI spatial corpus, which represents the spatial co-occurrence semantic relationship between different POI types. The Word2vec model is used to train ω m , change the POI type Converted to POI vector γ u , all POI vectors form a POI semantic space; for any target point tg of the virtual trajectory r , establish a radius of R * The buffer zone is composed of all POIs in the buffer zone as the target point tg r The POI context set represents the target point tg rThe semantic attribute information of POI is then converted into a POI vector set using the POI semantic space, and the average vector V of all vectors in the POI vector set is calculated. poi , taking it as the target point tg r The semantic attribute feature vector of .

[0079] Moreover, in step 5, the virtual trajectory timing operation feature vector V is concatenated by vector concatenation. op , the spatial attribute feature vector V of the virtual trajectory target point tile and the semantic attribute feature vector V poi Perform feature fusion:

[0080]

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

[0082] In chronological order, connect the spatiotemporal semantic fusion features of each time step to form a feature sequence of the virtual trajectory straj r represents the number of target points. The sequence contains the user's time, space and semantic information at each time step. The change of feature values ​​between sequences indicates the evolution of the user's temporal operation and spatial element preference. LSTM is used to build a temporal neural network model, and the feature time series FSeq is used as input data. The cross entropy loss is used as the loss function. The Adam optimizer is used to train the model for the virtual trajectory classification task. During the training process, two LSTM layers are used to process FSeq and capture its temporal dependencies. The hidden state and memory unit of each time step of the feature time series FSeq are calculated to form a hidden state time series H=<h(1),…,h(t),…,h(r)> As the output sequence of the LSTM layer, extract the hidden state h(r) of the last time step and enter 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 represents the weight, b g Indicates bias.

[0085] The present invention also provides a virtual trajectory classification system of spatiotemporal semantic fusion, which is used to implement the virtual trajectory classification method of spatiotemporal semantic fusion as described above.

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

[0087] Alternatively, it comprises a readable storage medium, wherein the readable storage medium stores a computer program, and when the computer program is executed, the virtual trajectory classification method of spatiotemporal semantic fusion as described above is implemented.

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

[0089] 1) Based on local dynamic sorting, the virtual trajectory time series noise is reduced to improve the quality of virtual trajectory data; 2) The spatiotemporal semantic expression of the virtual trajectory is defined, and the virtual trajectory is simplified into two parts: temporal operation and target point, to ensure the spatiotemporal continuity of the user access process; 3) A temporal operation representation model is proposed to extract the temporal operation features, a ground feature color template is constructed to extract the spatial attribute features of the target point, and a POI semantic space is established to extract the semantic attribute features of the target point; 4) The temporal, spatial and semantic features are integrated by using a temporal neural network model to capture the temporal dependency of the virtual trajectory and realize the accurate classification of the virtual trajectory; 5) The scheme of the present invention is simple and convenient to implement, and has strong practicality. It solves the problems of low practicality and inconvenience in practical application in related technologies, can improve user experience, and has important market value. BRIEF DESCRIPTION OF THE DRAWINGS

[0090] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

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

[0092] Figure 2 This is an example diagram of virtual track timing denoising using a local dynamic sorting algorithm according to an embodiment of the present invention.

[0093] Figure 3 It is the composition relationship between the timing operation and the target point in the virtual track of the embodiment of the present invention. DETAILED DESCRIPTION

[0094] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in combination with the drawings and embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0095] Example 1

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

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

[0098] The spatial interaction process of users in PMSP forms a virtual trajectory, which records the spatiotemporal process of user operations and access to vector map tiles. There are delays in both the loading of vector map tiles and the generation of access logs, which leads to uncertainty in the trajectory points in a small-scale temporal range, which is specifically manifested as the frequent cross-layer jumps of the hierarchical attributes of the virtual trajectory, which violates the operation rules of the map and forms the temporal noise of the virtual trajectory. The present invention collects massive user access logs in PMSP, constructs virtual trajectories based on the temporal relationship of access tile records in user access logs, reorders the virtual trajectories based on local dynamic sorting, reduces the frequent cross-layer jumps in the virtual trajectory, and realizes virtual trajectory temporal noise reduction.

[0099] Step 1.1, virtual trajectory construction.

[0100] The vector map tiles visited by each user in the public map service platform are extracted to construct a virtual track. The central coordinates of the vector map tiles represent the spatial position of the track point, and each track point is described by longitude, latitude, level, and time.

[0101] The vector map tiles visited by each user in PMSP are extracted to construct a virtual trajectory. Each vector map tile represents a trajectory point, and the center coordinates of the vector map tile represent the spatial position of the trajectory point. Therefore, the virtual trajectory traj is defined as:

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

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

[0104] Step 1.2, denoise the virtual trajectory timing based on the local dynamic sorting algorithm.

[0105] The virtual trajectory is traversed by moving the window, the movement cost and potential difference cost are calculated, the sorting cost function is minimized, and the dynamic sorting adjustment of the trajectory points is realized.

[0106] The timing noise of virtual trajectories is essentially the local incorrect sorting of trajectory points in timing. The goal of the local dynamic sorting algorithm is to use a local window to scan the virtual trajectory sequence, and to reduce the frequent jumps of the virtual trajectory hierarchical attributes through dynamic sorting adjustment of the trajectory points in the window, making the virtual trajectory sequence smoother.

[0107] Since the temporal noise of the virtual trajectory has local characteristics, the sorting process needs to control the implementation range. Therefore, the virtual trajectory traj is traversed with 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 Represents the i-th trajectory point of the virtual trajectory traj, which is also the starting point of the window.

[0110] Then, the cost of the sorted virtual trajectory within the 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 trajectory point p i is the sorting cost of the window at the starting point, M(i,j) represents any trajectory point p in the window j and starting point p i The movement cost caused by swapping positions, L(i,j) represents the potential difference cost of the window.

[0113] The calculation method of M(i,j) is:

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

[0115] wherein, i and j are the subscripts of the trajectory points p i and p j .

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

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

[0118]

[0119] wherein, i and j are the subscripts of the trajectory points p i and p j , 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 exchange positions.

[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] wherein, 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 achieves the optimal sorting and the virtual trajectory time series noise reduction by seeking the balance between the two. Figure 2 For the example of the local dynamic sorting algorithm for virtual trajectory time series noise reduction, Figure 2In (a), the solid line represents the level change of the original virtual trajectory, and the dotted line represents the level change of the virtual trajectory after local dynamic sorting. The solid line has frequent small fluctuations, while the dotted line is relatively smoother. Figure 2 (b) shows the relationship between LDS and window length k. As k increases, LDS tends to be stable. When k is 7, LDS is the smallest. In this embodiment, k is 10.

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

[0125] In PMSP, users browse maps by panning the same layer and zooming in and out across layers. Maps at different levels can provide different browsing perspectives. Low-level vector map tiles have low resolution and larger spatial range, while high-level vector map tiles have high resolution but small spatial range. Previous studies have shown that zooming in indicates increased visit interest, while zooming out indicates decreased visit interest. That is, when users find potential browsing targets in the map, they tend to zoom in to display high-resolution detailed information, forming target points in virtual trajectories. Target points represent the set of trajectory points with maximized visit interest. From the perspective of the user's browsing process, the virtual trajectory can be simplified as a target point transfer process achieved through temporal operations (zooming in, zooming out, panning, etc.), which includes two parts: temporal operations and target points:

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

[0127] Where straj represents the virtual trajectory after local dynamic sorting, and the timing operation op i Represents the user operation process in the virtual trajectory, the target point tg i represents the spatial semantic features of user preferences in the virtual trajectory, and r represents the number of target points. Figure 3 The composition relationship between the timing operations and the target points in the virtual trajectory is demonstrated, and the two alternately form the virtual trajectory.

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

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

[0130]

[0131] In the formula, op i Indicates timing operation, Pan miIt indicates translation operation on the same layer, Zoommi indicates cross-layer operation, and the length d of operations at different timings is relatively random and generally different.

[0132] The same-layer panning operation refers to the user moving or dragging the map window in the same layer to obtain vector map tiles in the new field of view, thereby expanding the browsing range and finding potential browsing targets, that is, the next target point. Therefore, the same-layer panning operation can be described by the moving distance and moving direction, which are defined as follows:

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

[0134] Where Tiledis represents the tile distance moved during the timing operation, and Direction represents the moving direction, which is quantified in four directions: east, south, west, and north.

[0135] Virtual trajectories are three-dimensional trajectories, and there are scale differences between trajectory points in different layers. The Euclidean space distance between adjacent trajectory points at a low level is generally much larger than that between adjacent trajectory points at a high level, that is, the traditional spatial distance measurement method cannot avoid the scale difference problem under different layers. When users browse PMSP, vector map tiles are basic spatial units, and vector map tiles are pictures of fixed pixel size. They are invariant in different levels and are only related to the user's sequential operation process. Therefore, the present invention uses tile distance to calculate the user's moving distance to avoid scale problems between different levels. 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 Respectively represent the column number and row number of the vector map tile p, col q and row q Respectively represent the column number and row number of the vector map tile q.

[0138] Cross-layer operation refers to the zooming in and out of the map. Due to the scale differences between layers, the spatial areas of adjacent layers in cross-layer operation must be contained. Therefore, there is no moving distance and moving direction, only level changes, that is:

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

[0140] In the formula, zoom represents the level of zooming in or out, and in the actual virtual trajectory, Zoom mi May span multiple levels.

[0141] Pan mi and Zoom mi These are two mutually exclusive basic operations. In order to unify the two, the present invention further defines an operation distance Opdis, which represents the number of sequential operations required from one target point to the next target point. 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 accommodate when the 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 levels of enlargement or reduction |zoom|. When moving on 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 above two.

[0145] Through the above definition of timing operations, the present invention constructs a timing operation sequence Op from a virtual track:

[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}, r represents the number of target points.

[0148] Compared with the basic operation in formula (8), op i Added extended attributes.

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

[0150] The target point is a track point with semantic information in the virtual track. It has similar connotations to the stop points of the GPS track. Compared with other track points, the target point has stronger user access interest. Although the target point only has spatial position coordinates, it provides an index for the spatial attribute features such as associated POIs and spatial elements, which is the key to analyzing user access interests. The present invention uses the HGMM (Hierarchical Gaussian Mixture Model) model to extract the target points in the virtual track, and then construct the target point sequence Tg:

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

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

[0153] Step 3: construct a timing operation representation model and extract the timing operation feature vector.

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

[0155] Users in different fields have different demands for map data at different levels, which leads to different temporal operation behaviors. For example, in the field of vehicle navigation, users' temporal operations are mainly manifested as high-level translation, while in the field of river (lake) chief system, spatial element query and browsing are the main demands. Users tend to repeatedly zoom in and out to query the locations of spatial elements such as rivers and lakes. Therefore, mining the temporal operation characteristics of users in different fields is the key to achieving accurate classification of users in different fields. The temporal operation Op represents the operation process of transferring from one target point to another. It has the characteristics of irregular temporal length and high data dimension, which makes temporal operation modeling difficult. The present invention constructs a temporal operation representation model, converts irregular and high-dimensional temporal operations into dense vectors of fixed dimensions through deep learning, and learns the potential distribution law of user operation behaviors, thereby supporting the task of classifying users in different fields.

[0156] The temporal operation representation model adopts the Encoder&Decoder structure, which is an unsupervised deep learning architecture suitable for unlabeled data and widely used for data dimensionality reduction and feature extraction. In the model design process, the temporal relationship of temporal operation attributes such as layers, zooming in or out levels, moving directions, tile distances, and operation distances is the key to feature extraction. Therefore, the basic unit in the Encoder&Decoder structure selects the temporal model LSTM (Long Short Term Memory).

[0157] The Encoder & Decoder structure is a symmetrical deep learning structure. The Encoder module converts the input sequence into a dense vector of fixed dimension, and then the Decoder module decodes the dense vector into the input sequence. The model convergence is achieved through repeated learning, where the dense vector is the time series operation feature vector. In the present invention, the Encoder module uses a double-layer LSTM to model the time series features. LSTM is the basic time series modeling unit. The time series operation sequence Op is input into the Encoder module, and the h output by LSTM is converted into the time series feature vector. end Calculate the distribution parameter mean μ and variance σ of the time series operation feature vector:

[0158]

[0159] Where W μ With W σ represents the weight, b μ With b σ Denotes the deviation, h end 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 by the mean μ and variance σ:

[0161] V op =μ+σε ε~N(0,1) (16)

[0162] Where ε represents the parameter of the 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, is the initial state of the Decoder module, Op' is the output of the Decoder module, tanh and sigm are activation functions, is the state of LSTM in the Decoder module, trained using the Adam optimizer, and iteratively updates the weight W z , W out and deviation b z 、b out .

[0166] The time series operation representation model is actually the application of the variational autoencoder model in the field of time series modeling. The loss function loss includes 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, and is used to calculate the distance between the probability distribution N(μ,σ) of the time series operation feature vector and the normal distribution N(0,1), that is:

[0169]

[0170] The MSE loss function requires that the reconstruction error between the input data Op of the Encoder module and the output data Op' of the Decoder module be minimized, that is:

[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 point.

[0174] Vector map tiles are converted from vector maps and contain line and surface features. They are frequently accessed and widely used. The virtual track target points are matched with vector map tiles and POIs to extract spatial elements and characterize the spatial attribute features and semantic attribute features 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 and the vector map tile are overlaid for analysis, and the vector map tiles that overlap with the target point in space are selected. A terrain color template is constructed to extract the proportion of terrain element categories in the vector map tiles to form a spatial attribute feature vector to characterize 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 track target point and extracts the spatial attribute characteristics of the target point. Vector map tiles contain line and surface element information, and different spatial attributes of ground objects are marked by color. Compared with remote sensing tiles, it is easier and more accurate to extract spatial attributes through the color of vector map tiles.

[0178] The present invention extracts the spatial attributes of vector map tiles by constructing a feature color template. The feature color template describes the RGB color values ​​of different feature types at different levels. The template is formed by collecting vector map tiles of different feature elements and identifying the RGB color values ​​of the feature elements. Generally speaking, the feature color templates of different public map service platforms are not the same. The present invention takes Tiandi Map as the research object. According to the feature classification standard of Tiandi Map, the template includes five major feature categories of roads, water, green space, land, and residential areas, as well as 10 minor feature categories. Considering the problem of hierarchical display of minor feature categories in the map synthesis process, for example, the functional area only appears at the [14, 15, 16, 17, 18] level, therefore, the extraction range of different minor feature categories is also divided by layer. The template is marked with RGB. Considering the subtle differences in the colors of vector map tiles, an error interval of ±2 is adopted. The specific template is shown in Table 1.

[0179] Table 1. Color templates for vector map tiles in the world map (2023)

[0180]

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

[0182] V tile =[v1 … v e … v 10 ] (twenty one)

[0183]

[0184] Where V tile It is a 10-dimensional spatial attribute feature vector, where each dimension represents a small category of land features; Indicates the number of pixels of this feature class in the tile; v e Indicates the pixel ratio of this ground feature subclass.

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

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

[0187] 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 to characterize the semantic attribute characteristics of the virtual trajectory target point.

[0188] The semantic attributes of the virtual trajectory represent the category attributes of the user's visit area corresponding to the trajectory point. The present invention uses POI spatial co-occurrence semantic representation. First, based on the spatial co-occurrence relationship of POI, the POI semantic space is constructed using POI data at a national scale to realize the vectorization of POI semantic attributes, and then the POI matched by the trajectory point is represented as a feature vector covering semantic attributes.

[0189] For any POI in the POI database covering POIs across the country, a buffer with a radius of R is established with the POI as the center, and the POI context set in the buffer is retrieved in Indicates the type of POI, and POIs are sorted from large to small based on the Euclidean distance. R and u are both hyperparameters and can be set in advance according to needs. In this embodiment, R is 1km and u is 30. m It is the POI spatial corpus, which represents the spatial co-occurrence semantic relationship between different POI types, such as the widely co-existing spatial proximity dependency between restaurants and shopping malls.

[0190] In order to achieve the quantitative expression of the co-occurrence semantic relationship of POI space, the Word2vec model is further used to train ω m , change the POI type Converted to POI vector γ u ,For any two POI vectors, the smaller the cosine similarity between them, the closer the vector distance is, and the higher the probability of co-occurrence of two types of POI space. All POI vectors form a POI semantic space.

[0191] For any target point tg of the virtual trajectory r , establish a radius of R * The buffer zone is composed of all POIs in the buffer zone as the target point tg r The POI context set represents the target point tg r The semantic attribute information of POI is then converted into a POI vector set using the POI semantic space, and the average vector V of all vectors in the POI vector set is calculated. poi , taking it as the target point tg r The semantic attribute feature vector of .

[0192] Step 5: construct a temporal neural network model to classify virtual trajectories with fusion of spatiotemporal semantic features.

[0193] The temporal operation feature vector in step 3, the spatial attribute feature vector in step 4.1, and the semantic attribute feature vector in step 4.2 are fused by vector concatenation, and a temporal neural network model is constructed to capture the temporal dependency of the virtual trajectory, classify and output the domain category of the virtual trajectory, and realize the domain classification of the virtual trajectory.

[0194] The temporal operation sequence Op of the virtual trajectory strak and the target point sequence Tg together constitute the user access process. The temporal operations, access vector map tiles and POIs of users in the same field have common characteristics, while different fields have differences. In this embodiment, agriculture, navigation, environmental protection, rivers and oceans with a large number of users in the world map are selected as classification fields.

[0195] The present invention extracts the virtual trajectory timing operation feature vector V in step 3 op In step 4, the spatial attribute feature vector V of the virtual trajectory target point is extracted tile and the semantic attribute feature vector V poi , the domain classification of virtual trajectories is realized by fusing three features. Since the three features are all vectors, the present invention performs feature fusion by vector splicing:

[0196]

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

[0198] In chronological order, connect the spatiotemporal semantic fusion features of each time step to form a feature sequence of the virtual trajectory straj r represents the number of target points. The sequence contains the user's temporal, spatial and semantic information at each time step. The change of feature values ​​between sequences indicates the evolution of the user's temporal operations and spatial element preferences. Therefore, extracting and learning the transfer relationship 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] The present invention uses the classic LSTM to construct a time series neural network model, takes the feature time series FSeq as input data, and trains the model to perform virtual trajectory classification tasks. During the training process, two LSTM layers are used to process FSeq and capture its time series dependencies, and the hidden state and memory unit of each time step (target point) of the feature time series FSeq are calculated to form a hidden state time series H=<h(1),…,h(t),…,h(r)> As the output sequence of the LSTM layer. Extract the hidden state h(r) of the last time step (target point) and enter 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 represents the weight, b g Indicates bias.

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

[0203] Example 2

[0204] Based on the same inventive concept, the present invention also provides a virtual trajectory classification system with 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 with 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, the virtual trajectory classification method with spatiotemporal semantic fusion as described above is implemented.

[0207] In specific implementation, the method proposed in the technical solution of the present invention can be implemented by technical personnel in this field using computer software technology to realize automatic operation process. System devices for implementing the method, such as computer-readable storage media storing the corresponding computer program of the technical solution of the present invention and computer equipment running the corresponding computer program, should also be within the scope of protection of the present invention.

[0208] The specific embodiments described herein are merely examples of the spirit of the present invention. Those skilled in the art may make various modifications or additions to the specific implementation cases described, or replace them with similar methods, without departing from the spirit of the present 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: The following steps are involved: Step 1, preprocessing the virtual trajectory timing based on local dynamic sorting; Step 1.1, extract the vector map tiles visited by each user in the public map service platform to construct a virtual track; Step 1.2, denoising the virtual trajectory time series based on a local dynamic sorting algorithm; Step 2, express the spatiotemporal semantics of the virtual trajectory by defining the timing operations and target points; Step 3: construct a temporal operation representation model and extract temporal operation feature vectors; Step 4, extracting the spatial semantic features of the virtual trajectory target point; Step 4.1, constructing a ground object color template to extract the spatial attribute feature vector of the target point; Step 4.2, establish POI semantic space to extract semantic attribute feature vectors of target points; In 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 with fusion of spatiotemporal and semantic features.

2. The virtual trajectory classification method of spatiotemporal semantic fusion according to 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 position of the trajectory point. Each trajectory point is described by longitude, latitude, level, and time. The virtual trajectory traj is defined as: through= <p1,p2,p3,…,p i ,…,p n > (1) In the formula, p i represents the i-th trajectory point, p i =(lon i ,lat i ,layer i ,time i ), 1≤i≤n, lon i Indicates longitude, lat i Indicates latitude, layer i Indicates level, time i represents time, and n is the number of trajectory points.

3. The virtual trajectory classification method of spatiotemporal semantic fusion according to claim 1, characterized in that: In step 1.2, the virtual trajectory traj is traversed with a window of length k, where the window w is defined as: w=<p i ,p i+1 ,…,p i+k-1 > (2) In the formula, p i represents the i-th trajectory point of the virtual trajectory traj; Then, the cost of the sorted virtual trajectory within the window relative to the original virtual trajectory is defined as: cost(i,j)=M(i,j)+L(i,j) (3) In the formula, cost(i,j) represents the cost of trajectory point p i is the sorting cost of the window at the starting point, M(i,j) represents any trajectory point p in the window j and starting point p i The movement cost caused by swapping positions, L(i,j) represents the potential difference cost of the window; The calculation method of M(i,j) is: M(i,j)=ji,(0 <i<j<k) (4) The calculation method of L(i,j) is: Where i and j are trajectory points p i 、p j The subscript of k is the window length, l j Represents the trajectory point p in the window j The potential difference cost L(i,j) represents the trajectory point p j and starting point p i The sum of the level differences between adjacent points in the window after swapping positions; The reordering process is expected to make the virtual trajectory sequence smooth by exchanging trajectory points in a local range, that is, to minimize the sorting cost. The virtual trajectory traj is traversed with window w, and the iterative cumulative sorting cost is calculated as: LDS(i)=LDS(i-1)+min{cost(i,j),(0 <i<j<k)} (6) 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 as the starting point, and LDS(i) represents the total sorting cost accumulated up to the i-th window position. When i = 1, LDS(1) = min(cost(1,j)).

4. The virtual trajectory classification method of spatiotemporal semantic fusion according to claim 1, characterized in that: In step 2, the virtual trajectory is simplified as a process of achieving target point transfer through timing operations, including two parts: timing operations and target points: strike={<op1,tg1> ,<op2,tg2> ,…, <op i ,tg i >,…, <op r ,tg r } (7) In the formula, straj represents the virtual trajectory after local dynamic sorting, the temporal operation op represents the user operation process in the virtual trajectory, and the target point tg i represents the spatial semantic features of user preferences in the virtual trajectory, and r represents the number of target points; The virtual track timing operation is defined from two perspectives: translation within the same layer and zooming in and out across layers, namely: In the formula, op i Indicates timing operation, Pan mi Indicates the same layer translation operation, Zoom mi Indicates cross-layer operation, d is the operation length; Furthermore, the translation operation on the same layer is described by the moving distance and moving direction, which are defined as follows: Pan mi ={Tiledis,Direction} (9) In the formula, Tiledis represents the tile distance moved during the timing operation, and Direction represents the moving direction, which is quantified in four directions: east, south, west, and north. The tile distance Tiledis is defined as: Tiledis=|col p -col q |+|row p -row q | (10) In the formula, col p and row p Respectively represent the column number and row number of the vector map tile p, col q and row q Respectively represent the column and row numbers of the vector map tile q; Cross-layer operations do not involve moving distance and direction, only layer changes, so Zoom mi Defined as: Zoom mi ={zoom} (11) In the formula, zoom indicates the level of enlargement or reduction; Pan mi and Zoom mi These are two mutually exclusive basic operations. In order to unify the two, the operation distance Opdis is further defined to represent the number of sequential operations required from one target point to the next target point. The specific calculation formula is as follows: In the formula, window represents the number of vector map tiles that the browsing window can accommodate when the user accesses the map; Based on the above definition of timing operations, the timing operation sequence Op constructed from the virtual trace is expressed as follows: On= <op1,op2,…,op i ,…,on r > (13) In the formula, op i ={layer i ,Tiledis i ,Direction i ,Zoom i ,Opdis i }, r represents the number of target points; Compared with the basic operation in formula (8), op i Added extended attributes; The HGMM model is used to extract the target points in the virtual trajectory, and then the target point sequence Tg is constructed: News= <tg1,tg2,…,tg i ,…,news r > (14) In the formula, tg i represents the i-th target point, which is composed of multiple trajectory points; r represents the number of target points.

5. The virtual trajectory classification method of spatiotemporal semantic fusion according to claim 4, characterized in that: In step 3, the temporal operation representation model adopts the Encoder&Decoder structure. The Encoder module converts the input sequence into a dense vector of fixed dimension, and then the Decoder module decodes the dense vector into the input sequence. After repeated learning, the model converges. The dense vector is the temporal operation feature vector. The Encoder module uses a double-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 the h output by LSTM is used. end Calculate the distribution parameter mean μ and variance σ of the time series operation feature vector: Where W μ With W σ represents the weight, b μ With b σ Denotes the deviation, h end represents the hidden state of the last time step; Based on the reparameterization technique, the time series operation feature vector V is calculated by the mean μ and variance σ op : V op =μ+σ,ε~N(0,1) (16) In the formula, ε represents the parameter that obeys the standard normal distribution; The Decoder module and the Encoder module have a symmetrical LSTM structure. The structure of the Decoder module is as follows: In the formula, is the initial state of the Decoder module, Op' is the output of the Decoder module, tanh and sigm are activation functions, is the state of LSTM in the Decoder module, trained using the Adam optimizer, and iteratively updates the weight W z , W out and deviation b z , b out ; The loss function loss includes two parts: KL divergence and mean square error, namely: loss=KL(N(μ,σ),N(0,1))+MSE(Op,Op') (18) KL divergence is used to calculate the distance between the probability distribution N(μ,σ) of the time series operation feature vector and the normal distribution N(0,1), that is: The MSE loss function requires that the reconstruction error between the input data Op of the Encoder module and the output data Op' of the Decoder module be minimized, that is: 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.

6. The virtual trajectory classification method of spatiotemporal semantic fusion according to claim 1, characterized in that: In step 4.1, the virtual trajectory target point and the vector map tile are overlaid and analyzed, the vector map tile that overlaps with the target point in space is selected, and a ground feature color template is constructed. The category of the ground feature element on the selected vector map tile is determined according to the RGB value, and the ground feature element category ratio of the vector map tile is extracted to form a spatial attribute feature vector to characterize 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 and identifying the RGB color values ​​of feature elements, a feature color template is formed. For any vector map tile, the extracted spatial attribute feature vector is: V tile =[v1…v e …v τ ] (21) Where V tile It is a τ-dimensional spatial attribute feature vector, where each dimension represents a small category of ground features; Indicates the number of pixels of this feature class in the tile; v e Indicates the pixel ratio of the small class of the feature.

7. The virtual trajectory classification method of spatiotemporal semantic fusion according to claim 1, characterized in that: In step 4.2, the POI context around the virtual trajectory target point is retrieved through the buffer, the average vector of the POI vector set is calculated, and a semantic attribute feature vector is formed to characterize the semantic attribute characteristics of the virtual trajectory target point. The specific operation is as follows: For any POI in the POI database covering POIs across the country, a buffer with a radius of R is established with the POI as the center, and the POI context set in the buffer is retrieved in Indicates the type of POI, and POIs are sorted from large to small according to the Euclidean distance. R and u are both hyperparameters, which can be set as needed. ω m It is the POI spatial corpus, which represents the spatial co-occurrence semantic relationship between different POI types. The Word2vec model is used to train ω m , change the POI type Converted to POI vector γ u , all POI vectors form a POI semantic space; for any target point tg of the virtual trajectory r , establish a radius of R * The buffer zone is composed of all POIs in the buffer zone as the target point tg r The POI context set represents the target point tg r The semantic attribute information of POI is then converted into a POI vector set using the POI semantic space, and the average vector V of all vectors in the POI vector set is calculated. poi , taking it as the target point tg r The semantic attribute feature vector of .

8. The virtual trajectory classification method of spatiotemporal semantic fusion according to claim 1, characterized in that: In step 5, the virtual trajectory timing operation feature vector V is concatenated by vector concatenation. op , the spatial attribute feature vector V of the virtual trajectory target point tile and the semantic attribute feature vector V poi Perform feature fusion: In the formula, represents the spatiotemporal semantic fusion features of trajectory straj at time step t, They represent the temporal operation feature vector, spatial attribute feature vector, and semantic attribute feature vector at time step t respectively; In chronological order, connect the spatiotemporal semantic fusion features of each time step to form a feature sequence of the virtual trajectory straj r represents the number of target points. The sequence contains the user's time, space and semantic information at each time step. The change of feature values ​​between sequences indicates the evolution of the user's temporal operation and spatial element preference. LSTM is used to build a temporal neural network model. The feature time series FSeq is used as input data, and the cross entropy loss is used as the loss function. The Adam optimizer is used to train the model for the virtual trajectory classification task. During the training process, two LSTM layers are used to process FSeq and capture its temporal dependency. The hidden state and memory unit of each time step of the feature time series FSeq are calculated to form a hidden state time series H=<h(1),…,h(t),…,h(r)> As the output sequence of the LSTM layer, extract the hidden state h(r) of the last time step and enter the fully connected layer to generate the classification prediction result: G(straj)=Softmax(W g h(r)+b g ) (24) In the formula, G(straj) represents the classification result of the virtual trajectory straj, Softmax represents the activation function, and W g represents the weight, b g Indicates bias.

9. A virtual trajectory classification system with spatiotemporal semantic fusion, characterized in that: The method comprises a processor and a memory, wherein the memory is used to store program instructions, and the processor is used to call the program instructions in the memory to execute a virtual trajectory classification method for spatiotemporal semantic fusion as described in any one of claims 1 to 8.

10. A virtual trajectory classification system with spatiotemporal semantic fusion, characterized in that: The method comprises a readable storage medium having a computer program stored thereon. When the computer program is executed, the method for classifying virtual trajectories with spatiotemporal semantic fusion according to any one of claims 1 to 8 is implemented.

Images