CN112749209B - Method for constructing mobile behavior patterns oriented to space-time data - Google Patents
Method for constructing mobile behavior patterns oriented to space-time data Download PDFInfo
- Publication number
- CN112749209B CN112749209B CN202011629525.3A CN202011629525A CN112749209B CN 112749209 B CN112749209 B CN 112749209B CN 202011629525 A CN202011629525 A CN 202011629525A CN 112749209 B CN112749209 B CN 112749209B
- Authority
- CN
- China
- Prior art keywords
- mobile behavior
- nodes
- node
- behavior pattern
- time
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
- G06F16/24—Querying
- G06F16/245—Query processing
- G06F16/2458—Special types of queries, e.g. statistical queries, fuzzy queries or distributed queries
- G06F16/2465—Query processing support for facilitating data mining operations in structured databases
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
- G06F16/28—Databases characterised by their database models, e.g. relational or object models
- G06F16/284—Relational databases
- G06F16/288—Entity relationship models
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
- G06F16/29—Geographical information databases
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/36—Creation of semantic tools, e.g. ontology or thesauri
- G06F16/367—Ontology
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Systems or methods specially adapted for specific business sectors, e.g. utilities or tourism
- G06Q50/01—Social networking
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Abstract
The invention belongs to the technical field of data mining, and particularly relates to a mobile behavior pattern representation method conforming to a user travel rule based on space-time data in a social network of a location service. The construction of the map mainly comprises three parts, wherein the first part is the construction of the mobile behavior map, and the first part carries out preprocessing operation on the sign-in data of the original user and then establishes the mobile behavior map; the second part is the construction of a gating graph neural network based on a mobile behavior graph, and the second part provides a method for fusing nodes and edges of the graph to node vectors and node update functions; the third part is the construction of a position prediction network based on a graph vectorized representation, which proposes a method of using updated node vectors for the position prediction network. The position prediction model based on the movement behavior pattern can cover the time-space data attribute from multiple angles and all directions, and can capture the movement characteristics of the user more accurately, so that the accuracy of the position prediction is improved.
Description
Technical Field
The invention belongs to the technical field of data mining, and particularly relates to a mobile behavior pattern representation method conforming to a user travel rule based on space-time data in a social network of a location service.
Background
With the rapid popularization and development of mobile intelligent devices, the mobile internet has penetrated into aspects of people's life. People can share and acquire information on a social network at any time and any place through an intelligent mobile terminal with a GPS, wiFi and other sensors. In many mobile terminal applications, such as take-away, taxi, restaurant and shopping, etc., users are required to provide location information, so a great amount of user data including space-time attributes is accumulated in a social network based on location services, and the data reflects abundant crowd movement information and is used in many research and application fields such as track mining, intelligent transportation and city calculation. The semantic information carried by the spatio-temporal data reflects the purpose of the user's activity in the moving process, namely the driving factors behind the user's movement between different places. Therefore, grasping and modeling the change rules of the activities can help us predict the position movement mode, and further optimize the position service and application, such as location recommendation, path planning, user behavior analysis, public health epidemic prevention and the like.
Position prediction is an important task of space-time data mining, and is a problem of judging possible positions accessed by a user in the future by capturing various movement behavior patterns and individual preferences through historical check-in records of the user. However, the multi-source heterogeneity, distributed sparsity, and complexity in individual movement patterns of spatio-temporal data make the effect of position prediction less than ideal. In the existing position prediction model, the problems of low utilization degree of time or space attribute and insufficient mining degree still exist, for example, data features cannot be fully reflected in the prediction model, or fusion degree among a plurality of features is low, so that a data structure input by the model cannot fully show a user movement rule, and the prediction effect is directly influenced. The deep utilization of spatiotemporal data attributes remains a significant problem that is not negligible.
The map representation of the data can show the direct or indirect association between the different attribute information of the data through the map structure, so that the space-time data is represented by the map, namely, a data organization structure with high connectivity is used as the input of a position prediction model, which is a feasible technical means. However, in the related achievements in the prior art, the problem that the attribute utilization is incomplete and the user movement characteristic representation is insufficient still exists in the spatio-temporal data under the graph representation. Therefore, the organization mode of the data features is further improved, and the organization of the space-time data aiming at comprehensively and thoroughly displaying the movement behaviors of the user has very important research significance.
In conclusion, the space-time data complex attributes are deeply mined and fused to form a reasonable and orderly data organization mode to improve the position prediction effect, so that the method is a reasonable and feasible innovation angle and has important research significance and application value.
Disclosure of Invention
The invention aims to provide a mobile behavior map construction method integrating multi-angle space-time attributes according to the actual requirements of a position prediction task in position service application and aiming at the characteristics of multi-source heterogeneity and extremely sparseness of space-time data, thereby realizing the deep utilization and mining of space-time data and finally achieving the purpose of improving the accuracy of predicting the future access position of a user.
The construction of the mobile behavior map aims at improving the prediction accuracy of the position prediction model through the whole coverage and representation of the time-space data attribute. The construction method of the map comprises three parts, namely, firstly, spatiotemporal data in a geographic social network are formed into map representation fused with multiple attributes; then taking the atlas as the input of the gating image neural network to perform the representation learning of the atlas node vector; finally, the user position prediction is realized by using the vectorized representation of the map.
In order to achieve the above purpose, the invention adopts the following technical scheme:
the method for constructing the mobile behavior pattern facing to the space-time data comprises the following steps:
s1, acquiring data such as user position information and the like, and preprocessing the data;
s2, constructing a mobile behavior map based on the processed user data;
and S3, placing the mobile behavior pattern into a gating map neural network and a position prediction network to realize position prediction.
According to the technical scheme, the data in the step S1 are obtained from a social network database based on the location service.
According to further optimization of the technical scheme, the data in the step S1 comprise user ID, place ID and time.
According to further optimization of the technical scheme, the node of the step S2 mobile behavior map adopts (t) i ，p i ) Representation, wherein t i For absolute time, p i Absolute position; the edges of the movement behavior patterns are directed edges, and the nodes in the front of the time slot point to the nodes in the back.
According to further optimization of the technical scheme, the edges of the mobile behavior atlas further comprise relative relations between every two nodes, each edge records relative time and relative position information between every two pairs of nodes, the relative time is represented by interval duration of time slots corresponding to the two nodes, and the relative position is represented by position distance corresponding to the two nodes.
According to further optimization of the technical scheme, the step S3 comprises the step of generating vectorized representation of the map by adopting a gating map neural network after the mobile behavior map is constructed.
According to further optimization of the technical scheme, vectorization of the map generated by the gating map neural network specifically comprises the following steps:
vector representation method of mobile behavior spectrum node in step S3.1
Definition 1: embedding vectors and vectors of mobile behavior pattern nodesRepresenting location information in a node, vector +.>Representing time slots in nodes, vector +.>Representing a geographic grid in a node, where d p 、d t And d r The dimensions of the three vectors, respectively, and thus the vector of the graph node v is represented as follows:
wherein the method comprises the steps ofRepresenting the connection of vectors, h v An embedded vector representing node v;
step S3.2 representing method of moving behavior pattern edge in gating pattern neural network
Definition 2: site transfer matrixWherein |V| is the number of nodes in the map, A tr For representing a transfer relationship between sites;
definition 3: time interval matrixA iv For representing time interval information between all node pairs associated by edges in a graph, for each node pair (v i ，v j ) The relative time correlation is expressed as:
wherein eta is E [0,1]As a preset parameter for controlling the decay rate of the correlation, |ts i -ts j I is the absolute time of two nodes ts i And ts j Is used for the time interval of (a),the larger the value, the higher the time correlation between two nodes;
definition 4: position distance matrixA dt The position distance information used for representing all the node pairs associated by the edges in the map is determined by a radial basis function, namely, the position relativity of each node pair can be represented by the value of the kernel function, and the position relativity can be specifically represented as:
wherein delta > 0 is used as a preset parameter for weighing the relative distance, c i And c j For node pairs (v i ，v j ) Is composed of longitude and latitude of the position of the node,the larger the value, the higher the position correlation degree between the two nodes is;
step 3.3 representing and learning method of mobile behavior pattern node vector in gating pattern neural network
Node set vector h v Site transfer matrix A tr Time interval matrix A iv And a position distance matrix A dt Fused into the round-robin function of the network in the following way:
the calculation describes a graph node set vector h v One update at time t-1, whereinFor the embedded vector of the ith node at the time t-1, b tr 、b iv And b dt For trainable bias parameters +.>U z 、/> U r W and U are trainable linear weight parameters, z v And r v Respectively used as an update gate and a reset gate in the gating graph neural network, sigma and Tanh are respectively the activation functions Sigmoid and Tanh in the network, and the node aggregate vector h is the dot multiplication operation of the matrix represented by the channel v After t times of updating the loop function, an embedded vector is formed +.>According to the technical scheme, further optimization is performed, attention of all nodes of the spectrum is firstly given to each node of the spectrum node vector updated by the circulation function, and then all nodes of the spectrum are accumulated through a linear neural network, and the method is specifically as follows:
wherein v is n For all embedded vector sets involving nodes in the graph, v i Is the ith node in the map, |V| is the number of nodes in the map, sigma is the Sigmoid activation function, W 1 、W 2 And b g For trainable neural network parameters, after a map node accumulated vector g is obtained, multiplying the map node accumulated vector g by all the place vectors and obtaining probability distribution of position prediction by using a Softmax function, specifically as follows:
the Cross-entopy is used as an objective function of the overall network, and is specifically as follows:
wherein k is the number of the true values corresponding to the mobile behavior pattern, namely the number of all follow-up visited places corresponding to the mobile behavior pattern;probability distribution for model predictors; y is i The independent heat vector is a true value.
According to the technical scheme, the values of the time interval matrix are determined through Newton's law of cooling.
According to the technical scheme, the time interval matrix definition is further optimized, and when the same two nodes exist in the same map and have multiple time intervals, the shortest time interval is reserved.
According to further optimization of the technical scheme, the preprocessed data in the step S1 comprises a user ID, a place ID, a time slot ID, a geographic grid ID, longitude and latitude.
The invention provides a brand new method for processing a position prediction task, namely, deducing a possible future visited place of a user by establishing a mobile behavior pattern of the user; the construction of the atlas realizes multi-azimuth and multi-angle utilization of complex attributes of the space-time data, comprehensively and deeply embodies the movement rule of the user, is beneficial to describing the movement behavior mode of the user more accurately, and further achieves the effect of improving the position prediction.
Compared with the prior art, the invention has the following advantages:
1) Compared with the prior art that the check-in records are input by a position prediction model in a discrete form or a sequence form, the method and the device utilize the graph structure to organize and represent space-time data, and can further improve the association representation among the user check-in records;
2) The invention respectively fuses the absolute and relative aspects of time and space into the representation of the map, and compared with the prior model, the attribute coverage aspect is more comprehensive, and the depicting degree of the user movement behavior is deeper and more accurate;
3) On the basis of the prior method for serializing the adjacent records, the invention further realizes the relation mining between the non-adjacent sign-in records, thereby more completely capturing the transfer modes of users between different positions;
4) The mobile behavior map modeling gating map neural network and the corresponding position prediction model are utilized to realize complete fusion of map elements, and the related experiments prove that the model improves the accuracy of position prediction.
Drawings
FIG. 1 is a diagram of a movement behavior pattern;
FIG. 2 is a frame diagram of a position prediction model based on a movement behavior pattern;
FIG. 3 is a flow chart of a method of position prediction based on a mobile behavior pattern;
FIG. 4 is an exemplary diagram of a user sign-in sequence conversion site transfer matrix;
FIG. 5 is a diagram of user data for two location service data sets, gowalla and Fourdage;
FIG. 6 is a graph showing a comparison of position predictions based on two data sets, gowalla and Fourda.
Detailed Description
In order to describe the technical content, constructional features, achieved objects and effects of the technical solution in detail, the following description is made in connection with the specific embodiments in conjunction with the accompanying drawings.
The construction of the mobile behavior map mainly comprises two parts, namely, constructing a map representation of the mobile behavior of a user by utilizing space-time data, referring to fig. 1, and a schematic diagram of the mobile behavior map is shown; and secondly, constructing a position prediction model based on the movement behavior pattern, wherein the model consists of a gating map neural network and a position prediction network, and the frame of the model is shown in fig. 2 and is a position prediction model frame diagram based on the movement behavior pattern. The steps from the collection of the space-time data to the generation of the final position prediction result are divided into 4 steps, a flow chart of which is shown in fig. 3, and is a processing flow chart of a position prediction method based on a mobile behavior pattern, and the implementation flow of the patent is further described below by combining a specific data set.
Step S1: preprocessing of location services data
1.1 data acquisition
The data in the geographic social networks Gowalla and Fourdan are used as network training and predicting objects, and the sign-in records of users are extracted to be used as basic data for constructing the mobile behavior patterns. Wherein the Gowalla data selects a public data set provided by Stanford university, which provides user check-in data from 2 months 2009 to 10 months 2010; the fourier data uses a data set provided by Dingqi Yang et al, relating to the united states region from 4 months 2012 to 9 months 2013. Referring to fig. 5, a diagram of user data for two location service data sets, golella and Foursquare, is shown. Both datasets employed the filtering and partitioning approach proposed by Yiding Liu et al, wherein for the golella dataset, the number of users accessing less than 15 and the number of places accessing less than 10 were filtered, the number of users finally obtained was 18737, the number of places was 32510, and the number of check-in records was 1278274. For the fourier dataset, the number of users accessing less than 10 places and the number of places accessing less than 10 people are filtered, the number of the finally obtained users is 24941, the number of places is 28593, and the number of check-in records is 1196248. The proportions of the training set, the validation set and the test set in the two data sets are 70%, 10% and 20%, respectively, and in order to guarantee the task requirement of predicting future access places, access records appear in the validation set and the test set, the places of the access records do not appear in the training set, and the access time is after the training set.
1.2 data processing
Since the movement behavior pattern is constructed in units of users, check-in records need to be grouped by user ID. The check-in time is divided into different time intervals, namely time slots according to the selected granularity, each specific time can be mapped into the corresponding time slot, and finally the specific time of the check-in is represented by the time slot ID. The position information of the check-in place is represented by a geographic grid, and the longitude and the latitude are respectively divided into a plurality of subintervals according to the longitude and latitude range related to all the check-in places in the acquired data, so that the geographic grid is formed. And mapping the longitude and latitude information of the place ID in the sign-in record to a corresponding grid, and finally, representing the position information by using the geographic grid ID. The attribute information related to the sign-in record after preprocessing comprises: user ID, location ID, time slot ID, geographic grid ID, longitude and latitude.
The method comprises the steps of dividing time slots and geographic grids according to the method in two data sets, namely dividing time into 24 x 7 = 168 time slots in a cycle, mapping specific access time in a user sign-in record into the time slots, and representing sign-in time information by corresponding time slot IDs; dividing longitude and latitude into 150 and 50 subintervals according to longitude and latitude ranges of the signing record positions in the data set, finally forming 7500 geographic grids, mapping the specific positions of the places in the signing record of the user to corresponding grids, and representing the position information by corresponding geographic grid IDs. Because the time spans of the sign-in records corresponding to the users are different, in order to facilitate the establishment of the patterns and the prediction models, all the sign-in records of the users are divided into a plurality of subintervals according to a certain time span, then in step S2, corresponding mobile behavior patterns are established for each interval, and through model debugging, the Gowallla pattern time span is finally determined to be 1 week, and the Fourdure pattern time span is finally determined to be 2 weeks.
Step S2: construction of movement behavior patterns
Each user corresponds to a set of behavioral patterns, each pattern of each user in the golella dataset is made up of 1 week check-in records, and each pattern is made up of 2 weeks check-in records for the fourier square dataset, according to the setup of step S1.
2.1 Mobile behavior atlas representation method integrating multiple space-time attributes
2.1.1 node representation method of mobile behavior patterns
Due to the time in the spatiotemporal dataAbsolute and relative concepts exist for space, respectively. Each node in the map represents a check-in place containing time and position absolute information, namely, the node information consists of three parts of information of a check-in place ID, a time slot ID and a position ID, as shown in a node expansion diagram on the right side of the map schematic diagram of fig. 1. For example, the user has uploaded two check-in records respectively, and the time slot ID and place ID information of record 1 is (t 1 ，p 1 ) Record 2 is (t 2 ，p 2 ) Wherein t is 1 And t 2 For absolute time, p 1 And p 2 Absolute position; and t is 1 And t 2 Is of relative time, p 1 And p is as follows 2 Is the relative position. The nodes of the mobile behavior patterns are used for representing absolute information in the space-time data, the location ID visited by the user in each node is used as basic information, and the nodes also comprise time slot IDs corresponding to the visit records and geographic grid information IDs. Thus each node represents a spatiotemporal semantic "place X located in a grid area at a time".
2.1.2 edge representation method of movement behavior atlas
The mobile behavior pattern is different from a sequence data structure which is used as a position prediction model input in the past, and is related to user sign-in records in a pattern mode, and every two nodes with access sequences in the pattern are related by edges. The edges of the graph are directed edges, and the nodes in the front of the time slot point to the nodes in the back. In addition, the edges of the map also comprise relative relations between every two nodes, as shown in the map schematic diagram of fig. 1, each edge records relative time and relative position information between every two pairs of nodes, wherein the relative time is represented by interval duration of corresponding time slots of two nodes; the relative position is expressed in terms of the position distance corresponding to the two nodes.
The edges of the atlas are simultaneously associated with adjacent and non-adjacent check-in records with a sequential access relationship, the direction of the edges is determined by the sequence of the check-in time of the two associated nodes, and the front node points to the rear node; each edge further contains relative time and location information of the associated node pair, i.e. the time interval and location distance of the node pair.
2.2 node vector representation learning method of movement behavior patterns
After the mobile behavior pattern is constructed, a position prediction model is utilized to generate a prediction result, and a frame diagram of the model is shown in fig. 2. The map is taken as input, the vector representation is carried out by adopting a gating map neural network, then the nodes are updated, and finally, the result is output through a position prediction network, and the whole processing flow is shown in figure 3. Therefore, in order to obtain the map node vector through network training and learning, the neural network structure of the map needs to be constructed according to the following flow:
2.2.1 vector representation method of map nodes
Definition 1: the embedding vector of the mobile behavior pattern node is defined as follows:
(Vector)representing location information in the node; vector->Representing a time slot in a node; vector quantityRepresenting a geographic grid in a node, where d p 、d t And d r The dimensions of the three vectors, respectively. The vector of map nodes v is thus represented as follows:
wherein the method comprises the steps ofRepresenting the connection of vectors, h v Representing the embedded vector of node v.
2.2.2 representing method of map edge in gating map neural network
Definition 2: site transfer matrixWherein |V| is the number of nodes in the map. A is that tr For representing the transfer relationship between sites. The matrix consists of two adjacent sub-matrices, the values in the two sub-matrices represent the connection weights of the edges, the directional relation between the nodes of the map is described, and the values are determined by the condition of going out and going in between the nodes connected by the edges, and are consistent with the transfer behaviors of users in different places in reality. To sign in sequence [ V 1 ，V 4 ，V 2 ，V 3 ，V 4 ，V 2 ]For example, a site transfer matrix constructed by transfer relationships between nodes is shown in fig. 4.
Definition 3: time interval matrixA iv For representing time interval information between all "node pairs" in the graph that are associated by edges. The values of the matrix are determined by newton's law of cooling. In particular, an exponential decay function used to calculate the temperature of the target object in the law is converted into a correlation function between node pairs based on time intervals. For each node pair (v i ，v j ) The relative time correlation is expressed as:
wherein eta is E [0,1]As a preset parameter for controlling the decay rate of the correlation. Ts i -ts j I is the absolute time of two nodes ts i And ts j Is a time interval of (a) for a time period of (b).The larger the value, the higher the time correlation between the two nodes. When the same two nodes exist in the same map and a plurality of time intervals exist, the shortest time interval is reserved, so that the aim of maximizing the node correlation is fulfilled.
Definition 4: position distance matrixA dt For representing positional distance information between all pairs of nodes in the graph that are associated by edges. The value of the matrix is determined by a radial basis function, i.e. the position correlation of each node pair can be represented by the value of the kernel function, which can be expressed specifically as:
where δ > 0 is a preset parameter that trades off relative distance. c i And c j For node pairs (v i ，v j ) Is composed of longitude and latitude of the position of the node.The larger the value, the higher the position correlation between the two nodes.
2.2.3 method for learning representation of map node vectors in gated graph neural network
After nodes and edges of the mobile behavior patterns are defined by the methods of 2.1 and 2.2, the respective vector representations and matrix representations are put into a gating graph neural network and a position prediction network, and the node vectors are finally subjected to position prediction according to the mobile behavior patterns through model iterative training. Wherein the node set vector h v Site transfer matrix A tr Time interval matrix A iv And a position distance matrix A dt Fused into the round-robin function of the network in the following way:
the calculation describes a graph node set vector h v One update at time t-1, whereinThe embedded vector of the ith node at the time t-1; b tr 、b iv And b dt Is a trainable bias parameter; />U z 、/> U r W and U are trainable linear weight parameters; z v And r v Respectively serving as an update gate and a reset gate in the gating graph neural network, wherein sigma and Tanh are activation functions Sigmoid and Tanh in the network respectively; the ". As indicated above, represents the dot product of the matrix. After the node vector is updated by t times of cyclic functions, an embedded vector is formed>And then used in subsequent location prediction networks.
Step S3: construction of position prediction model based on gating map neural network
According to definition 1 in 2.2.1, the nodes of the mobile behavior pattern are formed by splicing 3 embedded vectors, namely a place vector, a time slot vector and a geographic grid vector. The edges and the information contained in the edges are described by 3 types of matrixes, namely a place transfer matrix, a time interval matrix and a position distance matrix, wherein the place transfer matrix describes the associated weight information of all edges in the map according to definition 2 in 2.2.2, and the matrix is formed by connecting the outgoing edge weight matrix and the incoming edge weight matrix of all nodes. According to definition 3 in 2.2.2, the time interval matrix describes the time interval information between all pairs of nodes with connected edges in the graph, the value of the matrix is determined by the interval duration of the time slots of the two nodes, wherein according to newton's law of cooling, the longer the interval time, the more the time correlation decays, wherein the decay parameter is set to 0.03. According to definition 4 in 2.2.2, the position distance matrix describes position distance information between all node pairs with edges connected in the map, the value of the matrix is determined by longitude and latitude information of two nodes, and according to a radial basis function, the closer the distance is, the higher the position correlation is, wherein the kernel function parameter is set to be 60. After the node vector and the side information matrix are set, updating the node vector information according to a cyclic function in the gating map neural network in 2.2.3, wherein the 3-class matrix is used as a channel for exchanging information between nodes and is introduced into a node updating process.
Step S4: generation of position prediction results
The updated map node vector of the circulation function firstly endows attention of all nodes of the map in each node, and then accumulates all nodes of the map through a linear neural network, which is as follows:
wherein v is n An embedded vector set for all related nodes in the map; v i Is the ith node in the map; the I and V are the number of nodes in the map; sigma is a Sigmoid activation function; w (W) 1 、W 2 And b g Is a trainable neural network parameter. After the map node accumulated vector g is obtained, multiplying the map node accumulated vector g by all the place vectors and obtaining the probability distribution of the position prediction by using a Softmax function, wherein the probability distribution is specifically as follows:
the Cross-entopy is used as an objective function of the overall network, and is specifically as follows:
wherein k is the number of the true values corresponding to the mobile behavior pattern, namely the number of all the follow-up visited places corresponding to the pattern;probability distribution for model predictors; y is i The independent heat vector is a true value.
In the prediction stage, the historical sign-in record of each user is firstly converted into one to a plurality of mobile behavior patterns; and then taking the map as the input of a gating map neural network, and finally obtaining the probability distribution of the place as a position prediction result through a circulation function and a position prediction network.
The updated node vectors are further used as the input of a position prediction network, attention information containing all nodes of the map is added to each map node vector according to the method, all map nodes are accumulated to be used as a map integral vector to be expressed, then the map integral vector is multiplied by all place vectors respectively, and finally the product is used as a position prediction result to generate probability distribution of places through a Softmax function.
In the training stage, for each sign-in record corresponding to each user, the sign-in time in the training data is represented in a single-hot vector mode by taking all recorded places after the map record is constructed as a true value. Thus one map corresponds to one to a plurality of true values. The global model uses Cross-entopy as an objective function. The position prediction model provided by the invention compares 6 position prediction models in Gowalla and Fourdwere data sets, and the prediction result is evaluated by 4 evaluation indexes of accuracy, recall, average accuracy mean value and normalized discount accumulated gain. Comparison of results as shown in fig. 6, each row corresponds to an evaluation index, where the left is the result of the golella dataset and the right is the result of the fourier dataset. The corresponding model of each result is shown in a legend, wherein the first six models respectively represent a comparison model MGMPFM, IRenMF, geoMF, rankGeoFM, geoPFM and SAE_NAD, and the last model represents a position prediction model based on a movement behavior pattern. The graph shows that the performance of the model provided by the invention is better than that of a comparison model, and the prediction effect of the invention on the position prediction task is improved.
It is noted that relational terms such as first and second, and the like are used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Moreover, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or terminal that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or terminal. Without further limitation, an element defined by the statement "comprising … …" or "comprising … …" does not exclude the presence of additional elements in a process, method, article or terminal device comprising the element. Further, herein, "greater than," "less than," "exceeding," and the like are understood to not include the present number; "above", "below", "within" and the like are understood to include this number.
While the embodiments have been described above, other variations and modifications will occur to those skilled in the art once the basic inventive concepts are known, and it is therefore intended that the foregoing description and drawings illustrate only embodiments of the invention and not limit the scope of the invention, and it is therefore intended that the invention not be limited to the specific embodiments described, but that the invention may be practiced with their equivalent structures or with their equivalent processes or with their use directly or indirectly in other related fields.
Claims (8)
1. A method for constructing a mobile behavior pattern oriented to space-time data is characterized by comprising the following steps:
s1, acquiring user position information data and preprocessing;
s2, constructing a mobile behavior map based on the processed user position information data;
s3, the mobile behavior map is put into a gating map neural network and a position prediction network to realize position prediction; step S3 comprises the steps of generating vectorized representation of the mobile behavior pattern by adopting a gating map neural network after the mobile behavior pattern is constructed;
the vectorization for generating the movement behavior map by the gating map neural network specifically comprises the following steps:
vector representation method of mobile behavior spectrum node in step S3.1
Definition 1: embedding vectors and vectors of mobile behavior pattern nodesRepresenting location information in a node, vectorRepresenting time slots in nodes, vector +.>Representing in a nodeGeography grid, where d p 、d t And d r The dimensions of the three vectors, respectively, and thus the vector of the mobile behavior pattern node set v is represented as follows:
wherein the method comprises the steps ofRepresenting the connection of vectors, h v An embedding vector representing a set v of mobile behavior atlas nodes;
step S3.2 representing method of moving behavior pattern edge in gating pattern neural network
Definition 2: site transfer matrixWherein |V| is the number of nodes in the mobile behavior pattern, A tr For representing a transfer relationship between sites;
definition 3: time interval matrixA iv For representing time interval information between all node pairs associated by edges in a movement behavior pattern, for each node pair (v i ,v j ) The relative time correlation is expressed as:
wherein eta is E [0,1]As a preset parameter for controlling the decay rate of the correlation, |ts i -ts j I is the absolute time of two nodes ts i And ts j Is used for the time interval of (a),the larger the value, the higher the time correlation between two nodes;
definition 4: position distance matrixA dt The value of the matrix is determined by a radial basis function, that is, the position correlation of each node pair can be represented by the value of the kernel function, which can be specifically expressed as:
wherein delta > 0 is used as a preset parameter for weighing the relative distance, c i And c j For node pairs (v i ,v j ) Is composed of longitude and latitude of the position of the node,the larger the value, the higher the position correlation degree between the two nodes is;
step 3.3 representing and learning method of mobile behavior pattern node vector in gating pattern neural network
Node set vector h v Site transfer matrix A tr Time interval matrix A iv And a position distance matrix A dt Fused into the loop function of the gated graph neural network as follows:
the calculation describes a mobile behavior pattern node set vector h v One update at time t-1, whereinFor the embedded vector of the ith node at the time t-1, b tr 、b iv And b dt For trainable bias parameters +.>U z 、U r W and U are trainable linear weight parameters, z v And r v Respectively used as an update gate and a reset gate in the gating graph neural network, sigma and Tanh are respectively the activation functions Sigmoid and Tanh in the gating graph neural network, and the node set vector h represents the dot multiplication operation of the matrix v After t times of updating the loop function, an embedded vector is formed +.>
2. The method for building a mobile behavior pattern based on spatiotemporal data according to claim 1, wherein the user location information data in step S1 is obtained from a social network database based on location services.
3. The method for building a mobile behavior pattern based on spatiotemporal data according to claim 1, wherein the user location information data in step S1 includes a user ID, a location ID, and a time.
4. The method for constructing a mobile behavior pattern for spatiotemporal data according to claim 1, wherein the nodes of the mobile behavior pattern in step S2 adopt (t) i ,p i ) Representation, wherein t i For absolute time, p i Absolute position; the edges of the movement behavior patterns are directed edges, and the nodes in the front of the time slot point to the nodes in the back.
5. The method for constructing a mobile behavior pattern according to claim 4, wherein the edges of the mobile behavior pattern further comprise a relative relationship between every two nodes, each edge records relative time and relative position information between every two pairs of nodes, wherein the relative time is represented by interval duration of time slots corresponding to two nodes, and the relative position is represented by position distance corresponding to two nodes.
6. The method for constructing a mobile behavior pattern based on spatiotemporal data according to claim 1, wherein the node vectors of the mobile behavior pattern updated by the cyclic function firstly give attention to all nodes of the mobile behavior pattern in each node, and then accumulate all nodes of the mobile behavior pattern through a linear neural network, specifically as follows:
wherein v is n For all embedded vector sets involving nodes in the mobile behavior pattern, v i Is the ith node in the mobile behavior spectrum, |V| is the number of nodes in the mobile behavior spectrum, sigma is a Sigmoid activation function, W 1 、W 2 And b g After the mobile behavior map node accumulated vector g is obtained for the trainable neural network parameters, multiplying the mobile behavior map node accumulated vector g by all the place vectors and obtaining the probability distribution of the position prediction by using a Softmax function, wherein the probability distribution is as follows:
the Cross-entopy is used as an objective function of the overall network, and is specifically as follows:
wherein k is the number of the true values corresponding to the mobile behavior pattern, namely the number of all follow-up visited places corresponding to the mobile behavior pattern;probability distribution for model predictors; y is i The independent heat vector is a true value.
7. The method for constructing a spatiotemporal data oriented mobile behavior map according to claim 1, wherein values of said time interval matrix are determined by newton's law of cooling.
8. The method for constructing a mobile behavior pattern according to claim 1, wherein the shortest time interval is reserved when there are multiple time intervals in the same mobile behavior pattern for the same two nodes in the time interval matrix definition.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011629525.3A CN112749209B (en) | 2020-12-31 | 2020-12-31 | Method for constructing mobile behavior patterns oriented to space-time data |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011629525.3A CN112749209B (en) | 2020-12-31 | 2020-12-31 | Method for constructing mobile behavior patterns oriented to space-time data |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112749209A CN112749209A (en) | 2021-05-04 |
| CN112749209B true CN112749209B (en) | 2023-08-29 |
Family
ID=75650743
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011629525.3A Active CN112749209B (en) | 2020-12-31 | 2020-12-31 | Method for constructing mobile behavior patterns oriented to space-time data |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112749209B (en) |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107194011A (en) * | 2017-06-23 | 2017-09-22 | 重庆邮电大学 | A kind of position prediction system and method based on social networks |
| CN111400620A (en) * | 2020-03-27 | 2020-07-10 | 东北大学 | User trajectory position prediction method based on space-time embedded Self-orientation |
| CN112153221A (en) * | 2020-09-16 | 2020-12-29 | 北京邮电大学 | Communication behavior identification method based on social network diagram calculation |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US10268646B2 (en) * | 2017-06-06 | 2019-04-23 | Facebook, Inc. | Tensor-based deep relevance model for search on online social networks |
-
2020
- 2020-12-31 CN CN202011629525.3A patent/CN112749209B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107194011A (en) * | 2017-06-23 | 2017-09-22 | 重庆邮电大学 | A kind of position prediction system and method based on social networks |
| CN111400620A (en) * | 2020-03-27 | 2020-07-10 | 东北大学 | User trajectory position prediction method based on space-time embedded Self-orientation |
| CN112153221A (en) * | 2020-09-16 | 2020-12-29 | 北京邮电大学 | Communication behavior identification method based on social network diagram calculation |
Non-Patent Citations (1)
| Title |
|---|
| 基于GGNN的栓母对知识图谱构建与应用;赵振兵等;《电网技术》;第[1]-[7]页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112749209A (en) | 2021-05-04 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Feng et al. | Learning to simulate human mobility | |
| CN110399565B (en) | Recursive neural network interest point recommendation method based on space-time period attention mechanism | |
| Shi et al. | Predicting origin-destination flow via multi-perspective graph convolutional network | |
| Hsieh et al. | Inferring air quality for station location recommendation based on urban big data | |
| CN110414732A (en) | A kind of trip Future Trajectory Prediction method, apparatus, storage medium and electronic equipment | |
| CN111783295B (en) | Dynamic identification and prediction evaluation method and system for urban community specific human behavior chain | |
| Zhao et al. | Incorporating spatio-temporal smoothness for air quality inference | |
| Dai et al. | Spatio-temporal deep learning framework for traffic speed forecasting in IoT | |
| CN111242352A (en) | Parking aggregation effect prediction method based on vehicle track | |
| CN116108984A (en) | Urban flow prediction method based on flow-POI causal relationship reasoning | |
| Vahidnia et al. | Modeling the spread of spatio-temporal phenomena through the incorporation of ANFIS and genetically controlled cellular automata: a case study on forest fire | |
| Huang et al. | Stpr: A personalized next point-of-interest recommendation model with spatio-temporal effects based on purpose ranking | |
| Yang et al. | Predicting and optimizing city-scale road traffic dynamics using trajectories of individual vehicles | |
| CN116884222B (en) | Short-time traffic flow prediction method for bayonet nodes | |
| Quan et al. | An optimized task assignment framework based on crowdsourcing knowledge graph and prediction | |
| CN113139140A (en) | Tourist attraction recommendation method based on space-time perception GRU and combined with user relation preference | |
| CN112749209B (en) | Method for constructing mobile behavior patterns oriented to space-time data | |
| Raju et al. | Micro-simulation of residential location choice and its variation | |
| CN112766240B (en) | Residual multi-graph convolution crowd distribution prediction method and system based on space-time relationship | |
| CN115510333A (en) | POI prediction method based on space-time perception and combined with local and global preferences | |
| Wang et al. | A hypergraph-based hybrid graph convolutional network for intracity human activity intensity prediction and geographic relationship interpretation | |
| Hussain et al. | Accurate Location Prediction of Social-Users Using mHMM. | |
| CN113971496A (en) | Urban traffic network state evolution trend prediction method and system under influence of activities | |
| Luan et al. | An efficient target tracking approach through mobile crowdsensing | |
| Yuan et al. | A novel learning approach for citywide crowd flow prediction |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |