Method and system for extracting access interest topics in public map service platform
By collecting user access logs, building POI semantic spaces in the public map service platform, and using SVD models, extracting user access interest topic vectors, the problem of inaccurate user access interest extraction in the existing technology is solved, and more accurate user interests are achieved Analytics and personalized services.
Patent Information
- Application Number
- CN202210091398.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-01-26
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2042-01-26
AI Technical Summary
In the existing public map service platform, the accuracy of user access interest extraction is limited by ignoring the spatial semantic characteristics between spatial elements. The traditional theme model only considers the frequency of user accessing spatial elements rather than co-occurrence, resulting in inaccurate interest extraction.
By collecting user access logs, defining user access trajectories, reducing dimensionality into point feature trajectories, building POI semantic space, establishing user-POI matrix, using SVD model matrix to extract topic vectors, and mapping them to POI semantic space to express user access interests.
It improves the accuracy of user access interests and the gain of spatial semantic information, simplifies the complexity of user access trajectory, and supports the mining of access interests and personalized geographic information services.
Smart Images

Figure CN114491313B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of cyberspace information services, and in particular to a solution for extracting access interest topics in a public map service platform. Background Art
[0002] With the deep integration of geographic information and the internet, geographic information services have become an indispensable part of people's daily lives, providing users with spatial information services such as coordinate positioning, target query, and route planning. As a fundamental public service platform, the Public Map Service Platform (PMSP) provides comprehensive, massive, and dense spatial information. However, this overly dense information increases users' information filtering and screening costs, as well as the platform's processing costs, reducing the quality of geographic information services. Mining user behavior and understanding semantics within the PMSP to help users proactively discover geographic information of interest is key to addressing these issues and creating an intelligent and user-friendly PMSP.
[0003] During the browsing and visiting process, users are driven by the purpose of their visit. The tiles they browse on the map will revolve around targets of interest. For example, when users search for nearby natural scenic spots, their browsing targets may include mountains, rivers, lakes, islands, etc. The content visited by users is the above POI (Point of Interest). How to extract the POIs visited by users from the tiles is the key to constructing the representation of visit interests. The key to mining visit interests lies in the co-occurrence of spatial elements visited by users. The distribution of spatial elements is affected by urban regions, local customs, functional types, etc., and there are inherent spatial association patterns. For example, there are usually restaurants around shopping malls. How to enrich the expression of visit interests by combining the spatial proximity distribution patterns of different types of spatial elements is the key to improving the accuracy of visit interest mining. Mining common visit interests based on a series of complex spatial elements visited by each user is the core issue of characterizing the composition of visit interest preferences.
[0004] Existing research has confirmed that visitor interest is closely related to spatial elements, such as densely populated areas, major roads, coastlines, and points of interest (POIs) (natural landscapes, amusement parks, sports fields, museums, zoos, etc.). Furthermore, models such as least squares regression and artificial neural networks have been proposed to quantify the impact of spatial elements on user visit behavior. However, this research characterizes visitor interest based on the frequency of users visiting spatial elements, failing to consider the common patterns of visitor interest, which can be represented by the co-occurrence of users visiting different spatial elements.
[0005] The process of mining visit interests is similar to the topic extraction problem in natural language processing. Traditional topic models extract core words that express the main idea from text, while visit interest mining extracts key spatial elements (with high co-occurrence probability) from the spatial elements visited by each user as expressions of visit interests. Previous research has combined the classic topic model LDA (Latent Dirichlet Allocation) with the one-dimensional text content of user visits in web services to extract user visit interests. However, in PMSP, user visit content is tiles and the spatial elements surrounding the tiles. Therefore, constructing a topic model based on the two-dimensional spatial distribution of spatial elements is key to mining visit interests. Previous research on interest extraction based on the LDA model has only considered the co-occurrence frequency between the spatial elements visited by users, ignoring the inherent spatial correlation patterns of spatial elements, which limits the accuracy of visit interest extraction. Summary of the Invention
[0006] Existing technologies use statistics on tile access frequencies to represent access interests and use traditional topic models such as LDA to extract access interest topics, but ignore the spatial semantic features between spatial elements, which limits the accuracy of access interest extraction.
[0007] To achieve the above-mentioned purpose, the technical solution proposed in the present invention is a method for extracting access interest topics in a public map service platform, comprising the following steps:
[0008] S1. Collect massive user access logs accumulated in the public map service platform PMSP. Based on the temporal relationship of tile access records in the user access logs, define user access trajectories of surface elements. Based on local spatial density sampling, convert them into point element user access trajectories to achieve dimensionality reduction of user access trajectories.
[0009] S2. Establish a POI spatial corpus based on the spatial proximity of massive POIs, construct a POI semantic space, and achieve POI quantification;
[0010] S3: Match the user access trajectory points obtained in S1 with the nearest neighboring POIs, segment the user access trajectory by user access session, and combine the POI quantitative expression in the POI semantic space obtained in S2 to construct a user-POI matrix to represent the mapping relationship between users and visited POIs;
[0011] S4: Decompose the user-POI matrix obtained in S3 based on the SVD model matrix, extract the topic vector, map the topic vector to the POI semantic space in S2, match the nearest neighbor TopK POIs as the expression of the user's visit topic interest, and output the corresponding geographic information.
[0012] Moreover, the definition of the user access trajectory in step S1 is as follows,
[0013] In PMSP, users access the map through a browsing window. The user access trajectory is composed of the movement of the browser web page or the access window in the mobile phone, which constitutes a continuous surface element trajectory. Based on the temporal relationship of the user access process, the user access trajectory PMSPtraj= <pa1,pa2,pa3,…,pa p ,…,pa M >,
[0014] in,
[0015] pa p =({<col1,row1> ,<col2,row2> ,…, <col q ,row q >},layer p ,t p ),0≤p≤
[0016] M, M refers to the length of the user's access trajectory, pa p is the pth trajectory point in the user's access trajectory;
[0017] pa p It is a plane of trajectory PMSPtraj, layer p Representation hierarchy;
[0018] {<col1,row1> ,<col2,row2> ,…, <col q ,row q >} represents the tiles contained in PMSPtraj,
[0019] col1,col2,…,col q Indicates the row number of the tile, row1, row2, ..., row q Indicates the column number of the tile.
[0020] Moreover, the dimensionality reduction of the user access trajectory in step S1 is implemented as follows:
[0021] S11, time division, including dividing the user's access trajectory into sessions using a preset time λ as a threshold to distinguish multiple access behaviors of the user within a day;
[0022] S12, scale division, includes segmenting the user access trajectory based on hierarchical attributes for each session to avoid changes in the spatial granularity between trajectory points caused by hierarchical changes;
[0023] S13. Constructing a spatiotemporal cube of the user's access trajectory, including simulating the user's access window with a 3x3 template in the spatial dimension and correlating the data of the previous second with the data of the next second in the temporal dimension to form a spatiotemporal cube;
[0024] S14, spatial density calculation, including calculating the local density value in the template based on the space-time cube of the user's access trajectory;
[0025] S15, trajectory dimensionality reduction, including selecting the center point of the tile with the maximum density among all tiles at the current moment as the feature sampling point to achieve trajectory dimensionality reduction.
[0026] Moreover, the POI semantic space in step S2 is constructed as follows:
[0027] S21, POI spatial corpus construction, including converting the spatial distribution of POIs into text form, and all non-repeated types of POIs as a dictionary;
[0028] S22, POI semantic space construction, including training the POI space corpus in S21 based on the Word2vec model to construct the POI semantic space.
[0029] Furthermore, in step S3, the user's visit trajectory is expanded based on the POI to express the visit interest, which is achieved as follows:
[0030] S31, expanding the trajectory spatial context based on the POI, including traversing all user-accessed trajectory points in each session, establishing a spatial buffer, retrieving several nearest POIs, and using the POI type as the spatial context semantic information of each trajectory point;
[0031] S32. Constructing a user-POI matrix, including combining the vectors of POI types in the POI semantic space in S22 to quantize the POI space context into a matrix; averaging the POI vectors in the same access session to obtain an average POI vector representation of the access session, and constructing an average POI matrix of the user access session based on all sessions.
[0032] Furthermore, the method for extracting the access interest topics based on the topic model in step S4 is as follows:
[0033] S41, matrix decomposition based on the SVD model, including expressing the user's interests as a user-POI matrix A, which is composed of user access sessions and POI vectors, and then performing matrix decomposition on the user dimension of the matrix A based on the SVD model to obtain an interest-POI matrix;
[0034] S42. Thematic interest extraction, including representing each vector of the interest-POI matrix as a thematic vector of the access interest, mapping the thematic vector to the semantic space of the POI, approximately expressing it through the existing POI, and calculating the distance D(v,P) between each interest v and all POI types P in the interest-POI matrix based on the cosine similarity of the vector, where D(v,P)∈[-1,1], the larger |D(v,P)| is, the more similar P is to the interest v, and selecting several most similar POIs of the nearest neighbors as the description of the access interest.
[0035] On the other hand, the present invention also provides a system for extracting access interest topics in a public map service platform, which is used to implement the method for extracting access interest topics in a public map service platform as described above.
[0036] Furthermore, the following modules are included,
[0037] The first module is used to collect massive user access logs accumulated in the public map service platform PMSP. Based on the temporal relationship of tile access records in the user access logs, it defines the user access trajectory of surface elements. Based on local spatial density sampling, it converts them into point element user access trajectories to achieve dimensionality reduction of user access trajectories.
[0038] The second module is used to establish a POI spatial corpus based on the spatial proximity of massive POIs, construct a POI semantic space, and realize POI quantification;
[0039] The third module is used to match user access trajectory points with the nearest neighboring POIs, segment the user access trajectory based on user access sessions, and construct a user-POI matrix based on the quantitative expression of POIs in the POI semantic space to represent the mapping relationship between users and visited POIs;
[0040] The fourth module is used to decompose the user-POI matrix based on the SVD model matrix, extract the topic vector, map the topic vector to the POI semantic space, match the nearest neighbor TopK POIs as the expression of the user's visit topic interest, and output the corresponding geographic information.
[0041] Alternatively, it includes a processor and a memory, the memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute the above-mentioned method for extracting access interest topics in a public map service platform.
[0042] Alternatively, it includes a readable storage medium having a computer program stored thereon, and when the computer program is executed, it implements the above-mentioned method for extracting access interest topics in a public map service platform.
[0043] The beneficial effects of the present invention are as follows: in response to the potential expression of visit interests in PMSP and the difficulty in accurately quantifying and expressing them, a multi-dimensional user visit trajectory is defined based on the user zooming-in and zooming-out operation process; a user visit trajectory dimensionality reduction method is proposed to simplify the user visit trajectory of surface elements to the user visit trajectory of point elements, thereby reducing the complexity of the user visit trajectory; the spatial attribute expression of the user visit trajectory is expanded based on the nearest neighbor POI, and the inherent distribution pattern of POI is constructed in the POI semantic space representation combined with China's massive POI data, thereby achieving spatial semantic information gain of visit interests; a visit interest topic model is proposed based on the spatial co-occurrence of POIs, and the visit interest composition expressed based on POIs is mined.
[0044] 1. Establish user access trajectories to represent the user access process. Based on local spatial density sampling of user access trajectories, reduce the dimension of user access trajectories, simplify the complexity of user access trajectories, and support access interest extraction and mining.
[0045] In existing research, scholars usually define access interests based on the frequency of user access to tiles, ignoring the spatiotemporal continuity of the user access process. This paper defines user access trajectories based on the user access process, restores the user's movement process of accessing PMSP from log data, and establishes user access trajectories.
[0046] The multidimensional nature of continuous surface element trajectories generated by user visits in PMSPs poses challenges for modeling and managing user access trajectories. To simplify the representation of user access trajectories, this paper proposes a dimensionality reduction method for user access trajectories. While taking into account the spatiotemporal continuity of trajectories, a spatiotemporal cube of user access trajectories is constructed to achieve dimensionality reduction from surface element trajectories to point element trajectories based on maximizing local density.
[0047] 2. Establish a POI semantic space based on the spatial co-occurrence of POIs to achieve quantification of discretely distributed POIs.
[0048] POIs are important spatial elements that characterize visit interests. Matching POIs with user visit trajectories can materialize visit interests. However, discretely distributed POIs are difficult to incorporate into model calculations, and quantifying them is a key challenge of this invention. By analogy with text processing in natural language processing, POIs are compared to characters, and their spatial distribution is converted into sentences. A POI corpus is then established, and the Word2vec model is used to vectorize POIs. This essentially characterizes the spatial co-occurrence between POIs, i.e., the inherent distribution pattern of POIs, while preserving the semantic information of POI distribution.
[0049] 3. Combine the matrix decomposition method to extract topic vectors and map the POI semantic space to parse and access interest topics.
[0050] The present invention realizes the extraction of access interest topic vectors in the user-POI matrix based on SVD model matrix decomposition. The essence is to mine the topic vectors of clustered topics. The topic vectors also come from the POI semantic space, so the topic vectors are mapped to the semantic space, and the topic vectors are analyzed based on the POIs of the nearest neighbors of the topic vectors to realize the expression of access interest topics and solve the problem of lack of interpretability of the topic vectors.
[0051] The solution of the present invention is simple and convenient to implement and has strong practicality. It solves the problems of low practicality and inconvenience in actual application existing in related technologies, can improve user experience, and has important market value. BRIEF DESCRIPTION OF THE DRAWINGS
[0052] Figure 1 This is a flowchart for extracting topics of interest accessed in a public map service platform according to an embodiment of the present invention.
[0053] Figure 2 Schematic diagram of the process of forming a user access trajectory in the same layer according to an embodiment of the present invention.
[0054] Figure 3 This is a schematic diagram of reducing the dimension of a surface element trajectory to a point element trajectory according to an embodiment of the present invention.
[0055] Figure 4 Schematic diagram of the space-time cube and spatial density calculation of user access trajectories according to an embodiment of the present invention.
[0056] Figure 5 This is a schematic diagram comparing user access trajectories before and after dimensionality reduction according to an embodiment of the present invention.
[0057] Figure 6 This is a schematic diagram of a POI context constructed based on the nearest neighbor method according to an embodiment of the present invention.
[0058] Figure 7 This is a schematic diagram of an example of a POI semantic space according to an embodiment of the present invention. DETAILED DESCRIPTION
[0059] The technical solution of the present invention is described in detail below with reference to the accompanying drawings and embodiments.
[0060] like Figure 1 As shown, an embodiment of the present invention provides a method for extracting access interest topics in a public map service platform, comprising the following steps:
[0061] S1. Definition of user access trajectories and trajectory dimensionality reduction: We collect massive user access logs accumulated in the Public Map Service Platform (PMSP). Based on the temporal relationship between tile access records in the user access logs, we define user access trajectories for surface features. These trajectories are then converted into point feature user access trajectories based on local spatial density sampling to achieve dimensionality reduction of user access trajectories.
[0062] In PMSP, users access the web map through a browser or mobile application. Each service request generated by the access corresponds to the tile covered by the browsing window. The user access trajectory is the plane composed of the visited tiles, and as the browsing window moves, more tiles are continuously loaded to form a continuous plane. Therefore, the user access trajectory is composed of continuous surface elements, that is, a multi-dimensional continuous surface element user access trajectory, such as Figure 2 shown.
[0063] In order to characterize the user access process, we first define the user access trajectory:
[0064] PMSPtraj= <pa1,pa2,pa3,…,pa p ,…,pa M > (1)
[0065] Among them p =({<col1,row1> ,<col2,row2> ,…, <col q ,row q >},layer p ,t p ), 0≤p≤M, M refers to the length of the user's access trajectory, pa p is the pth trajectory point in the user's access trajectory. pa p It is a plane of trajectory PMSPtraj, layer p Indicates a hierarchy.<col1,row1> ,<col2,row2> ,…, <col q ,row q >} represents the tiles contained in PMSPtraj, col1, col2, ..., col q Indicates the row number of the tile, row1, row2, ..., row q Indicates the column number of the tile.
[0066] User access trajectories are continuous surface feature trajectories, which have higher dimensions than continuous point feature GPS trajectories. Furthermore, the spatial extents of feature trajectories at different time points overlap, and the time and space of the trajectories are a one-to-many mapping. This makes it difficult to distinguish between the trajectory processes and hinders modeling and management of user access trajectories. Converting continuous surface feature trajectories into continuous point feature trajectories is a key issue in dimensionality reduction.
[0067] Figure 3 The following diagram shows how to reduce the dimensionality of continuous surface feature trajectories into continuous point feature trajectories. User access trajectory dimensionality reduction starts with the user access trajectory generation process. Trajectories are segmented in time and space, taking into account temporal and spatial continuity to construct a user access trajectory spatiotemporal cube. Based on the user access trajectory spatiotemporal cube, the local density maximum is calculated. The point with the highest local density within the current time is found as the sampling point for the current time. Finally, the trajectory is reconstructed according to the time series to obtain a complete trajectory. The specific process is as follows:
[0068] (1) Time division: The user access trajectory is divided into sessions using the time λ as the threshold to distinguish the user's multiple access behaviors within a day, where λ is an empirical value determined by the specific PMSP.
[0069] Users typically visit a PMSP multiple times a day, and different visits should be divided into different sessions. Sessions can be considered independent user access trajectories. Session division is based on a time threshold λ, which is determined by the PMSP system. However, existing studies have shown that 10 minutes is a more appropriate threshold.
[0070] (2) Scale division: There are scale differences among user access trajectories at different levels. For each session, the user access trajectory is segmented based on the level attributes to avoid the change of spatial granularity between trajectory points caused by level changes.
[0071] (3) Constructing a spatiotemporal cube of user access trajectories: In order to eliminate the impact of network and system processing delays during user access to PMSP, a spatiotemporal cube of user access trajectories is constructed to simulate the window movement during user access. Usually, a browser window can load 3-4 tiles in the horizontal and vertical directions, so a 3x3 spatial template is constructed to simulate the user access window. Due to network or system delays, when a user triggers a map page, not all tiles in the window can be processed by the server at the same time. The tile triggering time recorded in the log will also be delayed. In order to eliminate the impact of delays and maintain the temporal continuity of the trajectory, the data of the previous second and the next second are associated in the time dimension to form a spatiotemporal cube of the user access trajectory, such as Figure 4 shown.
[0072] (4) Spatial density calculation: In order to reduce the dimension of the surface element trajectory to the point element trajectory, the point with the largest local density is calculated through the user access trajectory space-time cube as the feature sampling point to achieve trajectory dimension reduction. Using time as the index, the surface element trajectory of a trajectory at time t is recorded as pa t , associate the surface feature trajectory set at time t-1 and time t+1 {pa t-1 ,pa t ,pa t+1 Build a spatiotemporal cube of user access trajectories. Because tiles are distributed in a regular grid, local density values are calculated based on weighted templates. The weighted templates are used to distinguish tile access points recorded at the current moment and at the associated moment, emphasizing the degree of influence of different times on the current moment. To simplify the calculation, the weight of the current moment template is set to 2, and the weight of the associated moment template is set to 1. By scanning the spatial distribution area of the trajectory points, based on the spatiotemporal cube of the user access trajectory, the local density value λ(col, row, t) of each location is calculated as shown in Equation 2:
[0073]
[0074] Where n satisfies 0≤n≤9, which means the number of tiles covered by the template that are not empty, and w t The weight of the template can be empirically determined. col, row, and t represent the column, row, and time of the current tile, respectively. λ(col, row, t) includes the local density value one second before the current moment, the sum of the local density value at the current moment, and the local density value one second after the current moment.
[0075] (5) Trajectory dimensionality reduction: The center point of the tile Max(λ(col, row, t)) with the maximum density among all tiles at the current time t is selected as the simplified trajectory point. All simplified trajectory points are sorted in chronological order to obtain a unique trajectory point sequence for each user at each time. The simplified user access trajectory is then defined as Equation 3:
[0076] PMSPtraj= <pa1,pa2,pa3,…,pa p ,…,pa M > (3)
[0077] Among them p =(lon p ,lat p ,layer p ,t p ),0≤p≤M,pa p =(lon p ,lat p ,layer p ,t p) is a trajectory point that constitutes the PMSPtraj trajectory, lon p ,lat p , layer p Respectively represent the longitude, latitude and level of the space-time point, t p Indicates the moment value on the time series axis.
[0078] Figure 5 The comparison of user access trajectories before and after dimensionality reduction is shown, where part (a) is the user access trajectory before dimensionality reduction, and part (b) is the user access trajectory after dimensionality reduction. In this process, the density of the trajectory points in the neighborhood centered on the trajectory point is calculated. It is assumed that the greater the density of the point, the more it can represent the center of access within one second. The trajectory point with the maximum density is selected as the sampling point. The present invention performs dimensionality reduction with the smallest time unit and space unit, minimizes the information loss in the dimensionality reduction process, and solves the problems of user access trajectories being too high as continuous surface element trajectories, overlapping time and space distributions, and non-unique spatiotemporal mapping.
[0079] S2. POI semantic space construction: In this embodiment, a massive amount of POIs (Point of Interest) in various provinces and cities in China are used to establish a POI space corpus based on spatial proximity, and a POI semantic space is constructed based on the Word2vec model to achieve POI quantification.
[0080] S21. Construction of POI spatial corpus
[0081] In the field of natural language processing, a corpus refers to a large collection of structured text stored and processed electronically. A corpus consists of many documents, each of which is composed of sentences and words. The order of sentences and words in a document represents the context and relationships between the sentences and words. During corpus processing, a dictionary is generated based on all the words in the corpus. By analogy with the organization of text in natural language processing, the spatial distribution of POIs is converted into text to construct a POI spatial corpus. All non-repeating types of POIs constitute the dictionary. The core of this process lies in how to convert POIs in two-dimensional space into a one-dimensional sequence, taking into account the spatial location attributes of POIs.
[0082] Traverse all POIs in the POI library and create a POI document for each POI, which consists of (POI center ,POI context ), where POI center It refers to the traversed POI as the center coordinate, to establish a spatial buffer with a radius of R, and retrieve the k nearest POIs to form a POI context ={POI1, POI2,…, POI k}, where distance(POI center,POI o-1 ) <distance(POI center ,POI o ),1 <o≤k,POI o For POI context Any POI, POI context Sort by distance from near to far, distance(POI center ,POI o-1 ) indicates POI center With POI o-1 The Euclidean distance between Figure 6 R and k are both hyperparameters and can be set in advance as needed. R = 1km, k = 30.
[0083] S22. Constructing POI semantic space based on Word2vec model of CBOW structure
[0084] The present invention preferably trains the POI space corpus in S21 based on the Word2vec model to construct a POI semantic space. The Word2vec model uses the input training corpus to establish a neural network language model, maps each word to a feature vector based on the context, and retains the semantic information of the word in the generated vector. In the vector space, the closer the distance between two vectors, the higher the semantic similarity between the two words. The Word2vec model based on the CBOW structure uses surrounding words to predict the current word.
[0085] The size of the POI spatial corpus is H, POI h The context sampling window is c, and the maximum likelihood estimate of the model can be expressed as Equation 4:
[0086]
[0087] in Indicates POI h As the center, the POI context is constructed with c as the sampling window, which is a set of POIs around the central POI. h Not in this set. Defined in the CBOW model Formula 5:
[0088]
[0089] Where E is an energy function, E(POI a ,POI b )=-(POI a POI b ). Formula 5 shows that when the context is c, POI hProbability of co-occurrence. In the semantic space of POIs, each POI type can be represented as a vector. The closer the distance between two POI type vectors, the greater the probability of their co-occurrence. Figure 7 In this paper, POI vectors in POI semantic space are reduced to two-dimensional space based on TSNE, and then visualized.
[0090] S3. Expand user access trajectories based on POIs to express visit interests: Match user access trajectory points in S1 to the nearest neighboring POIs, segment user access trajectories based on user access sessions, and combine the quantitative expression of POIs in the POI semantic space in S2 to construct a user-POI matrix to represent the mapping relationship between users and visited POIs.
[0091] POIs are the main spatial elements that characterize user visit content. Extracting POIs covered by user visit trajectories is the key to characterizing visit interests. This paper converts user visit trajectories into POI sets and combines them with quantified POI vectors in the POI semantic space in S2 to quantify visit interests. Specifically, the following steps are involved:
[0092] S31. Extending trajectory spatial context based on POI
[0093] Similar to the method of constructing the POI corpus, all user access trajectory points in each session are traversed to establish a spatial buffer with a radius of R, and the k nearest POIs are retrieved. The type of POI is used as the spatial context semantic information of each trajectory point, that is, POI user ={POI1, POI2,…, POI s}, where POI user Represents the POI spatial context of the session user, POI s It indicates the specific POI type in the context, and s refers to the number of POIs visited by the user.
[0094] S32. Constructing a user-POI matrix
[0095] Combined with the vector of POI type in the POI semantic space in S22, the POI spatial context is quantized into a matrix:
[0096]
[0097] where |P 11 … P 1j | represents the vector of the first POI, j represents the dimension of the POI vector, and is also a hyperparameter of the Word2vec model. In this paper, a nationwide massive POI dataset is used, with j = 400.
[0098] Since the number of trajectory points in each session is different, in order to convert the POI and tile sets into a matrix of fixed size for subsequent model calculations, the POI vectors in the same access session are averaged to obtain the average POI vector representation of the access session:
[0099]
[0100] where 1 ≤ d ≤ s, and P1(d) represents the POI user any value in the first column vector of the matrix. Based on all sessions, an average POI matrix of user access sessions is constructed:
[0101]
[0102] where 1 < q < u, and u is the total number of sessions in the dataset.
[0103] S4. Extracting access interest topics based on the topic model: The user-POI matrix in S3 is decomposed based on the SVD (Singular Value Decomposition) model matrix to extract the topic vector, and the topic vector is mapped into the POI semantic space in S2, and the POIs of the nearest neighbor TopK are matched as the expression of the user's access topic interest, and the corresponding geographical information is output.
[0104] The user-POI matrix A in S3 represents the mapping relationship between the user and the POI. Further extracting the topic based on this matrix is the key to obtaining the access interest with co-occurrence. The specific method is as follows:
[0105] S41. Extracting the topic vector based on the SVD model matrix decomposition
[0106] The user's interest expression is the user-POI matrix A, which is composed of the user access session and the POI vector. In this invention, the user dimension of the A matrix is decomposed based on the SVD model:
[0107]
[0108] where A is a non-zero real matrix, A ∈ R u×j , the SVD decomposition has physical significance, U t represents the user-interest matrix, represents the interest-POI matrix, and ∑ t is a diagonal matrix composed of non-negative diagonal elements arranged in descending order. t represents the number of topic interests. During the matrix decomposition process, access interest clustering is achieved.
[0109] S42. Topic interest extraction
[0110] In S41 Each vector is represented as a topic vector of the access interest. The vector itself has semantic features. In order to materialize the vector, the topic vector is mapped to the semantic space of POI and expressed approximately by the existing POI. The cosine similarity calculation based on the vector The distance D(v,P) between each interest v and all POI types P in [-1,1] is calculated, where D(v,P)∈[-1,1]. The larger |D(v,P)| is, the more similar P is to interest v. The τ most similar POIs of the nearest neighbors are selected as the description of the visited interest:
[0111] Topic(v)={P1,P2,…,P τ},D(v,P1)>D(v,P2)>…>D(v,P τ ) (10)
[0112] Topic(v) is the user's visiting topic interest, which consists of multiple POI types P1, P2, ..., P τ τ is a hyperparameter, referring to the number of nearest neighbor POIs. In this embodiment, τ is preferably set to 5. Table 1 shows an example of access interest topics, where t = 10, i.e., 10 topics. Each topic consists of the top-5 nearest neighbor POIs. "0.915 * village-level place name" indicates that the distance between the POI "village-level place name" and the topic vector is 0.915. The larger the weight, the more representative the topic content.
[0113] Table 1 Examples of interview interest topics
[0114]
[0115] The present invention addresses the problem that it is difficult to quantify and extract the common access interests in the access process in PMSP. Based on the spatiotemporal continuity of user access behavior, the present invention proposes a user access trajectory, and based on the multi-dimensional attributes of the user access trajectory, proposes a dimensionality reduction method for the user access trajectory to simplify the user access trajectory; proposes a nearest neighbor method to realize the conversion of user access trajectory to POI; quantifies China's massive POIs based on the Word2vec model, constructs a POI semantic space that characterizes spatial proximity, and maps the POIs visited by users to the semantic space to realize the semantic information gain of the access interest space. A user-POI matrix is constructed in the semantic space, and topic vector extraction is realized based on SVD mode matrix decomposition. The access interests and their component expressions are mined in combination with the vector similarity in the POI semantic space. The present invention realizes the conversion of user access trajectory to access interests, clarifies the spatial distribution of access interests during the user's access to PMSP, helps PMSP service providers understand the access interests and access needs of massive users, and provides support for establishing user behavior portraits and providing personalized intelligent geographic information services.
[0116] In specific implementation, the method proposed in the technical solution of the present invention can be automatically run by those skilled in the art using computer software technology. System devices that implement the method, such as computer-readable storage media that store the corresponding computer program of the technical solution of the present invention and computer equipment that runs the corresponding computer program, should also be within the scope of protection of the present invention.
[0117] In some possible embodiments, a system for extracting topics of interest accessed in a public map service platform is provided, comprising the following modules:
[0118] The first module is used to collect massive user access logs accumulated in the public map service platform PMSP. Based on the temporal relationship of tile access records in the user access logs, it defines the user access trajectory of surface elements. Based on local spatial density sampling, it converts them into point element user access trajectories to achieve dimensionality reduction of user access trajectories.
[0119] The second module is used to establish a POI spatial corpus based on the spatial proximity of massive POIs, construct a POI semantic space, and realize POI quantification;
[0120] The third module is used to match user access trajectory points with the nearest neighboring POIs, segment the user access trajectory based on user access sessions, and construct a user-POI matrix based on the quantitative expression of POIs in the POI semantic space to represent the mapping relationship between users and visited POIs;
[0121] The fourth module is used to decompose the user-POI matrix based on the SVD model matrix, extract the topic vector, map the topic vector into the POI semantic space, and match the nearest neighbor TopK POIs as the expression of the user's visit topic interest.
[0122] In some possible embodiments, a system for extracting access interest topics in a public map service platform is provided, including 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 method for extracting access interest topics in a public map service platform as described above.
[0123] In some possible embodiments, a system for extracting access interest topics in a public map service platform is provided, including a readable storage medium having a computer program stored thereon. When the computer program is executed, a method for extracting access interest topics in a public map service platform as described above is implemented.
[0124] The specific embodiments described herein are merely illustrative of the spirit of the present invention. Persons skilled in the art may make various modifications, additions, or substitutions to the described specific embodiments without departing from the spirit of the present invention or exceeding the scope of the appended claims.
Claims
1. A method for extracting access interest topics in a public map service platform, characterized in that: The following steps are involved: S1. Collect massive user access logs accumulated in the public map service platform PMSP, define the user access trajectory of surface elements based on the temporal relationship of access tile records in the user access logs, and convert them into point element user access trajectories based on local spatial density sampling to achieve dimensionality reduction of user access trajectories; S2, establish a POI spatial corpus based on the spatial proximity of massive POIs, construct a POI semantic space, and realize POI quantification; S3, matching the user access trajectory points obtained in S1 with the nearest neighbor POIs, segmenting the user access trajectory by user access session, combining the POI quantitative expression in the POI semantic space obtained in S2, constructing a user-POI matrix to represent the mapping relationship between the user and the visited POI; S4, decompose the user-POI matrix obtained in S3 based on the SVD model matrix, extract the topic vector, map the topic vector to the POI semantic space in S2, match the nearest neighbor TopK POI as the expression of the user's access topic interest, and output the corresponding geographic information.
2. The method for extracting access interest topics in a public map service platform according to claim 1, characterized in that: The definition of the user access trajectory in step S1 is as follows: In PMSP, users access the map with a browsing window. The user access trajectory is composed of a continuous surface element trajectory consisting of the movement of the browser web page or the access window in the mobile phone. Based on the temporal relationship of the user access process, the user access trajectory PMSPtraj=<pa1,pa2,pa3,…,pa p ,…,pa M >, in, pa p =({<col1, row1>, <col2, row2>, ..., <col q , row q >}, layer p , t p ), 0≤p≤M, M refers to the length of the user's access trajectory, pa p is the pth trajectory point in the user's access trajectory; pa p is a plane of trajectory PMSPtraj, layer p Representation level; {<col1, row1>, <col2, row2>, ..., <col q , row q >} indicates the tiles contained in PMSPtraj, col1, col2, ..., col q Indicates the row number of the tile, row1, row2, ..., row q Indicates the column number of the tile.
3. The method for extracting access interest topics in a public map service platform according to claim 1, characterized in that: The dimensionality reduction of the user access trajectory in step S1 is implemented as follows: S11, time division, including dividing the user's access trajectory into sessions based on a preset time λ as a threshold, and distinguishing multiple access behaviors of the user within a day; S12, scale division, including for each session, segmenting the user access trajectory based on the hierarchical attributes to avoid the change of spatial granularity between trajectory points caused by the hierarchical change; S13, constructing a space-time cube of the user's access trajectory, including simulating the user's access window with a 3x3 template in the spatial dimension, and associating the data of the previous second and the next second in the time dimension to form a space-time cube; S14, spatial density calculation, including calculating the local density value in the template based on the space-time cube of the user's access trajectory; S15, trajectory dimension reduction, including selecting the center point of the tile with the maximum density among all tiles at the current moment as the feature sampling point to achieve trajectory dimension reduction.
4. The method for extracting access interest topics in a public map service platform according to claim 1, characterized in that: The POI semantic space is constructed in step S2 as follows: S21, POI spatial corpus construction, including converting the spatial distribution of POIs into text form, with all non-repeating types of POIs as dictionaries; S22, POI semantic space construction, including training the POI space corpus in S21 based on the Word2vec model to construct the POI semantic space.
5. The method for extracting access interest topics in a public map service platform according to claim 1, characterized in that: In step S3, the user's access trajectory is extended based on POI to express the access interest, which is implemented as follows: S31, expanding the trajectory spatial context based on the POI, including traversing all the user access trajectory points in each session, thereby establishing a spatial buffer, retrieving several nearest POIs, and using the type of the POI as the spatial context semantic information of each trajectory point; S32, constructing a user-POI matrix, including combining the vectors of POI types in the POI semantic space in S22, quantizing the POI space context into a matrix; averaging the POI vectors in the same access session to obtain an average POI vector representation of the access session, and constructing an average POI matrix of the user access session based on all sessions.
6. A method for extracting access interest topics in a public map service platform according to claim 1, 2, 3, 4 or 5, characterized in that: The implementation method of extracting the access interest topics based on the topic model in step S4 is as follows: S41, matrix decomposition based on the SVD model, including expressing the user's interest as a user-POI matrix A, which is composed of user access sessions and POI vectors, and then performing matrix decomposition on the user dimension of the matrix A based on the SVD model to obtain an interest-POI matrix; S42. Thematic interest extraction, including representing each vector of the interest-POI matrix as a thematic vector of the access interest, mapping the thematic vector into the semantic space of the POI, approximately expressing it through the existing POI, and calculating the distance D(v, P) between each interest v and all POI types P in the interest-POI matrix based on the cosine similarity of the vector, where D(v, P)∈[-1, 1], the larger the |D(v, P)|, the more similar P is to the interest v, and selecting several most similar POIs of the nearest neighbors as the description of the access interest.
7. A system for extracting access interest topics in a public map service platform, characterized in that: Used to implement a method for extracting access interest topics in a public map service platform as described in any one of claims 1-6.
8. The system for extracting access interest topics in a public map service platform according to claim 7, characterized in that: Includes the following modules, The first module is used to collect the massive user access logs accumulated in the public map service platform PMSP, define the user access trajectory of the surface element based on the temporal relationship of the tile access records in the user access log, and convert it into the user access trajectory of the point element based on local spatial density sampling to achieve dimensionality reduction of the user access trajectory; The second module is used to establish a POI spatial corpus based on the spatial proximity of massive POIs, construct a POI semantic space, and realize POI quantification; The third module is used to match the user's access trajectory points with the nearest neighbor POIs around them, segment the user's access trajectory based on the user's access session, and build a user-POI matrix based on the POI quantitative expression in the POI semantic space to represent the mapping relationship between the user and the visited POI; The fourth module is used to decompose the user-POI matrix based on the SVD model matrix, extract the topic vector, map the topic vector to the POI semantic space, match the nearest neighbor TopK POI as the expression of the user's access topic interest, and output the corresponding geographic information.
9. The system for extracting access interest topics in a public map service platform according to claim 7, characterized in that: It includes a processor and a memory, the memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute a method for extracting access interest topics in a public map service platform as described in any one of claims 1-6.
10. The system for extracting access interest topics in a public map service platform according to claim 7, characterized in that: It includes a readable storage medium, on which a computer program is stored. When the computer program is executed, it implements a method for extracting access interest topics in a public map service platform as described in any one of claims 1-6.
