CN113010803B - Prediction method for user access position in geographic sensitive dynamic social environment - Google Patents

Prediction method for user access position in geographic sensitive dynamic social environment Download PDF

Info

Publication number
CN113010803B
CN113010803B CN202110578856.7A CN202110578856A CN113010803B CN 113010803 B CN113010803 B CN 113010803B CN 202110578856 A CN202110578856 A CN 202110578856A CN 113010803 B CN113010803 B CN 113010803B
Authority
CN
China
Prior art keywords
user
access
social
time
neural network
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
Application number
CN202110578856.7A
Other languages
Chinese (zh)
Other versions
CN113010803A (en
Inventor
胥帅
许建秋
关东海
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Nanjing University of Aeronautics and Astronautics
Original Assignee
Nanjing University of Aeronautics and Astronautics
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Nanjing University of Aeronautics and Astronautics filed Critical Nanjing University of Aeronautics and Astronautics
Priority to CN202110578856.7A priority Critical patent/CN113010803B/en
Publication of CN113010803A publication Critical patent/CN113010803A/en
Application granted granted Critical
Publication of CN113010803B publication Critical patent/CN113010803B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/95—Retrieval from the web
    • G06F16/953—Querying, e.g. by the use of web search engines
    • G06F16/9536—Search customisation based on social or collaborative filtering
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/95—Retrieval from the web
    • G06F16/953—Querying, e.g. by the use of web search engines
    • G06F16/9537—Spatial or temporal dependent retrieval, e.g. spatiotemporal queries
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/047—Probabilistic or stochastic networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • 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

Abstract

The invention discloses a method for predicting a user access position in a geographic sensitive dynamic social environment, which comprises the following steps: firstly, grid division is carried out on the activity area of the mobile social network user; secondly, calculating the influence implicit expression of the social friends; thirdly, dynamically updating the implicit access preference of the user; then, regarding the position prediction problem as a multi-classification problem, and solving the probability distribution of the next access position of the user; constructing a loss function aiming at the constructed recurrent neural network model and carrying out model optimization to obtain an optimization parameter of the model; and finally, calculating the access probability of the next movement behavior of the user for different candidate positions according to the model, and sequencing according to the probability values in a descending order to obtain a prediction result list. The method gives consideration to geographical sensitivity and dynamic characteristics, and can effectively predict the next access position of the user in a geographical sensitive social environment.

Description

Prediction method for user access position in geographic sensitive dynamic social environment
Technical Field
The invention relates to a method for predicting a user access position in a geographic sensitive dynamic social environment.
Background
The mobile social network has information service characteristics of socialization, localization, mobility and the like, and mass position trajectory data generated by the mobile social network records the moving process of people in the real physical world, reflects the living and traveling habits of people and contains extremely rich space-time semantic information. The future access position of the user is predicted on the basis of analyzing and mining the position track of the user in the mobile social network, so that the method not only can bring values for activity recommendation and precise marketing of commodities, but also is beneficial to building intelligent traffic and intelligent cities, and therefore, the method has important practical and social meanings. The position service derived based on the user position prediction becomes an important support for the high-quality development of the digital industrialization in China, and is beneficial to promoting the integration of the information technology and the entity economy on a deeper level.
The social relationship is the inherent attribute of the mobile social network, the friend users often have consistent access preference, and the movement behaviors of the users are influenced by the friends explicitly or implicitly according to the social dependency theory, so that the predictability of the access positions of the users can be enhanced by reasonably utilizing the positions of the friends. In the prior art, the influence of friend relationships on future mobile behaviors of users is mostly quantified from a statistical angle, and the time sequence interactive behaviors of the users in a network are not considered, so that the prior art is difficult to be used for friend dynamic interactive modeling in a mobile social environment. Currently, exploration aiming at social factors is mainly oriented to an online socialized recommendation scene, and a friend interaction mechanism in a mobile social network is quite different from the socialized recommendation scene, and the fundamental difference is whether a user interacts with an article in the real world. In an online social recommendation (such as book and movie recommendation) scene, the online behaviors of people are not limited by region space, and the influence of social factors on the user behaviors can be clearly reflected.
However, the range of activities of people in the off-line physical space is limited by regions, even online friends with consistent interest and preference have difficulty in substantially influencing the movement behavior of the user, and the user has difficulty in accessing the positions visited by the online friends due to geographical distance. Relevant statistics indicate that there are no co-visited locations among more than 50% of social friends in the mobile social network, which means that it is difficult for traditional statistical methods and neural network models to deeply characterize social influence propagation in a mobile social environment.
Generally, in the prior art, social factors are regarded as static weights or fixed factors, and are integrated into a prediction model to improve prediction accuracy. However, as for social influence in a mobile social network, since the location tracks of friend users are logically interlaced with each other and are mutually stimulated as the tracks extend, the dynamic social influence has geographic sensitivity and dynamics, and the dynamically changing social influence is more difficult to model than the static social influence, so how to integrate the geographic sensitivity dynamic social influence to predict the future visiting location of the user is an important problem to be solved by the invention.
Disclosure of Invention
The invention aims to provide a method for predicting a user access position in a geographical sensitive dynamic social environment, which considers geographical sensitivity and dynamic characteristics and can effectively predict the next user access position in the geographical sensitive social environment.
In order to achieve the purpose, the invention adopts the following technical scheme:
a method for predicting a user access position in a geographic sensitive dynamic social environment comprises the following steps:
step 1, carrying out grid division on a user activity area of a mobile social network;
firstly, gridding a user activity area, dividing the user activity area into a plurality of square grid units with the same size, and expressing the geographic coordinates of the grid units by using the central coordinates of the square area;
the check-in positions in the check-in record set of the user correspond to the grid cells to which the user belongs one by one, all interest points in each grid cell are represented by the grid cell at the moment, and the check-in positions of the user are presented in a square grid form;
step 2, calculating the influence implicit expression of the social friends;
sequentially inputting the user sign-in record set into a recurrent neural network model according to a time sequence;
at the time t, calculating a geographic coefficient according to the geographic distance between the user and the friend, and performing weighted aggregation on the geographic coefficient and the hidden state of the friend by using a linear function to obtain linear aggregation of the hidden state of the social friend of the user at the time t;
then, activating the linear aggregation by utilizing a nonlinear function to obtain an implicit expression of the influence of the social friends at the moment;
step 3, dynamically updating the implicit access preference of the user;
dynamically updating the implicit access preference of the user by utilizing a hidden layer of the recurrent neural network at the moment of t + 1;
the updating process comprises the following steps: fusing the implicit access preference of the user at the time t, the implicit expression of the influence of social friends at the time t and the position vector of the user at the time t +1, and calculating the implicit access preference of the user at the time t +1 by using a recurrent neural network hidden layer;
step 4, carrying out model training on the recurrent neural network model;
considering the user access position prediction problem as a classification problem, and solving the probability distribution of the next access position of the user;
on the basis, an objective function of the recurrent neural network model is constructed, and model parameters of the recurrent neural network model are optimized by minimizing the objective function, so that the trained recurrent neural network model is obtained;
step 5, predicting the next access position of the user;
on the basis of the step 4, calculating the access probability of the user aiming at each candidate position by utilizing a trained recurrent neural network model for predicting the next access position of the user, and sequencing the access probabilities of all the candidate positions in a descending order;
and finally, selecting candidate positions corresponding to the top-K probability value to form an ordered list as a prediction result.
Preferably, step 2 is specifically:
defining t time user uiU with friendsjHas a geographical distance of li,jThe geographic coefficient r is calculated as followst i,j；
Figure GDA0003146530820000021
Wherein a represents a smoothing factor;
defining t time user uiFriend u ofjHas an implicit state of hj tUser uiThe hidden states of the friends are linearly aggregated according to the geographic coefficient to obtain the user uiLinear aggregation of hidden states of good friends at time t Hi t；
Figure GDA0003146530820000031
Wherein Ni is user uiA set of social friends of;
then, for Hi tCarrying out nonlinear activation to obtain the influence implicit expression a of the social friendsi t，
Figure GDA0003146530820000032
Wherein the function
Figure GDA0003146530820000033
Representing a non-linear activation function.
Preferably, step 3 is specifically:
at time t +1, the input variables of the recurrent neural network include not only the user uiPosition vector e at time t +1i vt+1And the method also comprises an influence implicit expression a of the social friends at the moment ti tAnd implicit Access preferences h of the user at time ti t；
Defining a user u at time t +1iHas an implicit access preference of hi t+1Then h isi t+1The update method of (1) is as follows:
Figure GDA0003146530820000034
wherein, the function GRU (-) represents a calculation unit of the gated recurrent neural network, and ^ represents the splicing operation.
Preferably, step 4 is specifically:
defining user uiSet of check-in records LiThe access position subsequence from the middle to the t-th moment is Li tAnd obtaining the probability distribution of the next visit position of the user by utilizing a softmax function, wherein the calculation process is as follows:
P(vi t+1＝vm|Li t)＝softmax(WT m·hi t)；
wherein v ismDenotes the m-th candidate position, vmE.g. V, V represents a position set;
vi t+1represents the user u at the time t +1iThe access location of (2);
P(vi t+1＝vm|Li t) The physical meaning of (A) is:
given user u before time tiAccess position sub-sequence L ofi tUser uiThe access position is v at time t +1mThe probability of (d);
w denotes a weight matrix for classification, WmRepresents the mth row of the weight matrix W;
constructing an objective function, wherein a negative log-likelihood Loss function Loss of all user check-in sets is expressed as:
Figure GDA0003146530820000035
wherein, P (v)i n+1|Li n) The physical meaning of (A) is: given user uiIs a subsequence L of the first n access positions ofi nUser uiIs vi n+1The probability of (d);
Lirepresenting user uiU represents a mobile social network user set;
and minimizing the Loss function Loss by using an adaptive moment estimation algorithm, completing autonomous step learning of gradient descent, obtaining model parameters of the recurrent neural network model, and obtaining the trained recurrent neural network model.
Preferably, step 5 is specifically:
step 5.1. for target user uiAnd arbitrary candidate position vo，voE to V, and calculating the target user u by utilizing the recurrent neural network model trained in the step 4iThe next access position is voAccess probability P (v)o|Li)；
Wherein, P (v)o|Li) Representing a given user uiSet of check-in records LiThe next access position of the user is voThe probability of (d);
step 5.2, sequencing the access probabilities of all the candidate positions in a descending order;
and 5.3, selecting candidate positions corresponding to the top-K probability value to form an ordered list as a final prediction result.
The invention has the following advantages:
as mentioned above, the invention relates to a method for predicting the user access position in a geographical sensitive dynamic social environment, the method carries out time sequence modeling aiming at the coupled sign-in behavior between friend users, and the hidden states of friends are aggregated through a novel social pooling structure at any time, so as to simulate the dynamic social interaction incentive in a real scene.
Drawings
FIG. 1 is a flowchart of a method for predicting a user's visiting location in a geo-sensitive dynamic social environment, according to an embodiment of the present invention;
FIG. 2 is a schematic diagram illustrating a user visit location prediction problem solved by the method of the present invention;
FIG. 3 is a schematic diagram of meshing according to an embodiment of the present invention;
FIG. 4 is a general block diagram of a method for predicting a user's visiting location in a geo-sensitive dynamic social environment according to an embodiment of the present invention.
Detailed Description
The invention is described in further detail below with reference to the following figures and detailed description:
the embodiment describes a method for predicting the user access position in a geographic sensitive dynamic social environment, so as to effectively predict the user access position in the geographic sensitive social environment.
Before describing the process of the present invention in detail, the following definitions or explanations are first given.
The check-in records of the user in the mobile social network are collected through a public interface program (API), and each check-in record consists of three elements, namely the user, the check-in time and the check-in position (i.e. the fun point).
On the basis of obtaining the user sign-in record set, the method and the system fuse the geographic distance and the social relationship, so that the implicit access preference of the user is determined by three factors, namely the self moving mode, the moving behavior of the friend and the geographic distance between the friend and the user.
Due to implicit access preferences among different users, the access positions of the next movement behaviors of a plurality of users can be predicted finally as the movement tracks are mutually influenced and updated step by step along with the extension of the movement tracks.
For convenience of explanation, the following definitions are given for the user visit location prediction problem in a geo-sensitive dynamic social environment:
defining U as a known set of mobile social network users, V as a set of locations, LiRepresenting user uiIn the set of check-in records, Ni denotes user uiThe set of social friends. For target user uiRequiring prediction of user u in the form of a top-K ordered listiNext visit location of, so that user uiThe location of the real visit is located as much as possible at the top of the prediction list.
The technical problem to be solved by the present invention can be illustrated by fig. 2. Suppose that the user 1, the user 2, the user 4 and the user 5 are social friends, and each user has a check-in record set of the user and reflects the life and travel habits of the user.
As for user 1, user 1 is known to be at t2、t3And t4Visiting and checking in to the school, restaurant and cafe in turn at the moment, and predicting the next movement behavior (i.e. t) of the user 15Time of day).
The access destination of the next movement behavior of the user 1 should consider not only the movement behavior pattern of the user 1 (as shown by solid arrows (r) and (c) in fig. 2), but also the movement behavior pattern of the social friend in a similar scene in the check-in record set:
assuming that the user 2 and the user 5 are closer to the user 1, the moving behaviors of the user 2 and the user 5 should have a larger influence on the access destination of the next moving behavior of the user 1 (as shown by solid arrows (c) and (c) in fig. 2);
assuming that the user 4 is far away from the user 1, the movement behavior of the user 4 should have a small influence on the visiting destination of the next movement behavior of the user 1 (as indicated by a dotted arrow (c) in fig. 2).
In a real situation, the movement behavior of each user at each moment can be influenced by the social friends, and the mutual influence of the movement behaviors among the friends is called dynamic social interaction stimulation in the embodiment of the invention.
The method and the system dynamically update the instant implicit access preference of the user based on the geographic distance between the mobile social network user and the friends and the implicit states of the friends, and realize the prediction of the next access position of the user in a classified mode on the basis.
Specifically, as shown in fig. 1, the prediction method includes the following steps:
step 1, considering that the shared check-in positions among friends are extremely sparse, most friends even have no interest points which are accessed together, and therefore, if the social relationship is directly utilized to predict the next access position of the user, the rationality is not good to a certain extent.
Based on this, the method of the present invention proposes to perform meshing on the user activity area, and use a square area with regular shape and suitable size as the access position of the user, as shown in fig. 3. The specific division process is as follows:
firstly, gridding a user activity area, dividing the user activity area into a plurality of square grid units with equal sizes, and expressing the geographic coordinates of the grid units by using the central coordinates of the square areas.
And the check-in positions in the check-in record set of the user correspond to the grid cells to which the user belongs one by one, all interest points in each grid cell are represented by the grid cell, and the check-in positions of the user are presented in a square grid form.
FIG. 3 is a schematic diagram of an exampleSuppose user u' visits location v in sequencea、vbAnd vcAfter the gridding processing of the user activity area in step 1, the position visited by the user u' is changed to A0,0、A0,0And A0,1。
By converting the position prediction problem of the interest point granularity into the position prediction problem of the area granularity, the rationality and the operability of predicting the future movement behavior of the user based on the social relationship can be improved.
FIG. 3 is a diagram illustrating the partitioning of a user activity area of a mobile social network, each grid cell Ai,jThe side length threshold value delta is set by considering both the prediction performance and the practical value, wherein the larger the side length threshold value delta is, the coarser the position granularity is, the higher the prediction precision aiming at the next access position of the user is, and the lower the corresponding practical value is; conversely, the smaller the side length threshold δ is, the finer the position granularity is, and the lower the prediction accuracy for the next access position of the user is, the higher the practical value corresponding thereto is.
On the basis of the above-mentioned grid division, the positions visited by the user (i.e. check-in positions) can be presented in the form of square areas in fig. 3. Under a real scene, a check-in position sequence formed by square grids with proper sizes among social friends overlaps in a certain proportion, access behaviors of the friends aiming at different interest points in the same region have similar space-time semantics to a great extent, and at the moment, the future access position of the user is predicted to have greater rationality based on social relations.
And 2, calculating the influence implicit expression of the social friends based on the geographic distance between the user and the friends.
To characterize the geo-sensitive dynamic social interaction incentives, the present invention proposes a novel recurrent neural network model, as shown in FIG. 4, that uses a recurrent neural network for time-sequential modeling of the check-in set for each user.
For each position v visited by a user, embedding the position v into a d-dimensional continuous vector space according to a neural embedding idea to obtain an embedded representation e of the position vvWherein e isv∈Rd，RdRepresenting a d-dimensional continuous vector space.
Each recurrent neural network can share information after operating for one step, and the hidden states of the friends (namely the hidden access preferences of the friends) at the moment are weighted and aggregated through a geographic sensitive social pooling structure, wherein the influence of the geographic distance between the user and the friends at the moment needs to be considered in the aggregation process.
For convenience of presentation, h is definedi tFor user uiImplicit Access preferences at time t, Hi tFor user uiLinear aggregation of the implicit states of social friends at time t, ai tFor social friends to user u at time tiIs used to represent the social influence of (a).
Under the network model shown in fig. 4, the next visiting location of each user will be determined by the user itself and the moving behavior of the friend together, and the closer the geographical distance between the user and the friend is, the greater the influence degree of the friend is.
In FIG. 4, to visually represent this geographically sensitive "Social pooling" structure, the process is represented using a dashed box and a Geo-Social influx Aggregation (abbreviated GS-I-A). In addition, the GRU unit is used as the operation unit of the recurrent neural network model, and the calculation efficiency is better than that of operation units such as LSTM.
User uiThe check-in record sets are sequentially input into the recurrent neural network model in time sequence.
Calculating a geographic coefficient according to the geographic distance between the user and the friend at each moment, and performing weighted aggregation on the geographic coefficient and the hidden state of the friend by using a linear function to obtain linear aggregation H of the hidden state of the social friend of the user at the moment ti t。
Then using nonlinear function to Hi tActivating to obtain the influence implicit expression a of the social friends at the momenti t。
The geographic coefficient values have the following meanings: the larger the geographic distance between the user and a friend at the time t is, the smaller the geographic coefficient between the two persons is, the smaller the influence degree of the sign-in behavior of the friend on the user is, and vice versa.
At time t, user u needs to be computediWith social friends ujGeographic coefficient of (u)jIs epsilon Ni. Let t time user uiU with friendsjHas a geographical distance of li,jGeographic coefficient of the moment rt i,jCalculated as follows:
Figure GDA0003146530820000061
where a denotes a smoothing factor. At time t, according to the recurrent neural network model shown in fig. 4, the hidden states of the social friends need to be linearly aggregated according to the geographic coefficient to obtain Hi tTo Hi tCarrying out nonlinear activation to obtain ai t。
Social friends ujHidden state h ofj tAccording to the geographic coefficient, carrying out polymerization to obtain Hi tThe calculation formula of (A) is as follows;
Figure GDA0003146530820000071
wherein Ni represents user uiThe set of social friends.
To Hi tCarrying out nonlinear activation to obtain a user u at the moment tiIs implicitly represented by the influence of social friends ofi tI.e. by
Figure GDA0003146530820000072
Function(s)
Figure GDA0003146530820000073
Representing a non-linear activation function, such as a linear modified unit activation function ReLU.
The process of step 2 can obtain that the method takes account of the geographical distance between the user and the friends and the implicit access preference of the friends of the user, and utilizes a novel geographical sensitive 'social pooling' structure to perform weighted aggregation and activation on the implicit access preference of the social friends, thereby obtaining the implicit expression of the influence of the social friends.
And 3, dynamically updating the implicit access preference of the user.
As shown in fig. 4, at the time t +1, the implicit access preference of the user at the previous time, that is, the time t, is dynamically updated by using the recurrent neural network hidden layer, so as to obtain the implicit access preference of the user at the time t + 1. The specific process is as follows:
at time t +1, the input variables of the recurrent neural network include not only the user uiPosition vector e at time t +1i vt+1And the method also comprises an influence implicit expression a of the social friends at the moment ti tAnd implicit Access preferences h of the user at time ti t。
ei vt+1Representing user uiVisit location v at time t +1t+1Embedded representation in d-dimensional vector space, ei vt+1∈Rd。
Defining a user u at time t +1iHas an implicit access preference of hi t+1，hi t+1Not only with user uiImplicit Access preference h at time ti tCorrelation, also influenced by social friends, then hi t+1The update method of (1) is as follows:
Figure GDA0003146530820000074
wherein the function GRU (-) represents a computational unit of a recurrent neural network in a gated form,
Figure GDA0003146530820000075
indicating a splicing operation.
It should be noted that the GRU (-) computation unit contains two input variables at each instant, i.e.
Figure GDA0003146530820000076
And hi tWhen the operation is carried out,
Figure GDA0003146530820000077
and hi tA splicing operation between two input variables is also required.
And 4, carrying out model training on the recurrent neural network model.
On the basis of the steps 1-3, the user visit location prediction problem is regarded as a classification problem.
Defining user uiSet of check-in records LiThe access position subsequence from the middle to the t-th moment is Li tAnd obtaining the probability distribution of the next visit position of the user by utilizing a softmax function, wherein the calculation process is as follows:
P(vi t+1＝vm|Li t)＝softmax(WT m·hi t)；
wherein v ismDenotes the m-th candidate position, vm∈V，vi t+1Represents the user u at the time t +1iThe access location of (2);
P(vi t+1＝vm|Li t) The physical meaning of (A) is:
given user u before time tiAccess position sub-sequence L ofi tUser uiThe access position is v at time t +1mThe probability of (d);
w denotes a weight matrix for classification, WmThe mth row of the weight matrix W is shown.
Aiming at the constructed recurrent neural network model with the social pooling structure, the method provided by the invention learns the model parameters of the recurrent neural network model by minimizing the negative log-likelihood loss. The specific process is as follows:
let user uiIs LiThen, an objective function of the recurrent neural network model is constructed, that is, a negative log-likelihood Loss function Loss of the check-in set of all the users is expressed as:
Figure GDA0003146530820000081
wherein, P (v)i n+1|Li n) The physical meaning of (A) is: given user uiIs a subsequence L of the first n access positions ofi nUser uiIs vi n+1The probability of (c).
Due to user uiSet of check-in locations LiKnowing, therefore, for each access position n, vi n+1Can be determined.
Wherein n belongs to [1, | Li|-1]，|Li| represents user uiSet of check-in records LiThe number of elements (c).
And minimizing the Loss function Loss by using an adaptive moment estimation algorithm, finishing autonomous step learning of gradient descent, obtaining model parameters of the recurrent neural network model, including a weight matrix W, and obtaining the trained recurrent neural network model.
Compared with the traditional recurrent neural network, the recurrent neural network model provided by the invention has the core difference that a plurality of recurrent neural networks (three recurrent neural networks shown as user 5, user 1 and user 2 in fig. 4) are coupled through a geographic sensitive social pooling structure, so that the model training stage needs to perform gradient joint back propagation aiming at a plurality of GRU units.
And 5, predicting the access position of the user by using the trained recurrent neural network model.
Step 5.1. for target user uiAnd arbitrary candidate position vo，voE to V, and calculating the target user u by utilizing the recurrent neural network model trained in the step 4iThe next access position is voAccess probability P (v)o|Li)。
Wherein, P (v)o|Li) Representing a given user uiSet of check-in records LiThe next access position of the user is voThe probability of (c).
P(vo|Li) By the formula
Figure GDA0003146530820000083
And (4) calculating. Wherein, WoRow o representing the weight matrix W;
Figure GDA0003146530820000082
representing user uiRecord L of signing iniThe last check-in of (c) corresponds to an implicit access preference.
Step 5.2, sequencing the access probabilities of all the candidate positions in a descending order;
and 5.3, selecting candidate positions corresponding to the top-K probability value to form an ordered list as a final prediction result.
As can be seen from the frame diagram of fig. 4, in the mobile social network, the location tracks of each user are coupled with each other, and their visiting locations at each time t are related and mutually influenced, so that the method of the present invention can predict not only the next visiting location of a certain user, but also the next visiting locations of a plurality of friends related to the certain user.
In conclusion, the method can simultaneously realize the prediction of the next access position of a plurality of users.
Taking the three users in FIG. 4 as an example, the three users are sequentially marked as u5、u1And u2. Assume their check-in record sets are L in order5、L1And L2According to the step 5, three users can simultaneously calculate the optional candidate positions v by utilizing the trained recurrent neural network modelpAccess probability P (v)p|L5)、P(vp|L1) And P (v)p|L2) And then, sequencing the access probabilities of all the candidate positions in a descending order, and selecting the candidate position corresponding to the top-K probability value to obtain a final prediction result.
It should be understood, however, that the description herein of specific embodiments is not intended to limit the invention to the particular forms disclosed, but on the contrary, the intention is to cover all modifications, equivalents, and alternatives falling within the spirit and scope of the invention as defined by the appended claims.

Claims (5)

1. A method for predicting a user's visiting location in a geo-sensitive dynamic social environment,
the method comprises the following steps:
step 1, carrying out grid division on a user activity area of a mobile social network;
firstly, gridding a user activity area, dividing the user activity area into a plurality of square grid units with the same size, and expressing the geographic coordinates of the grid units by using the central coordinates of the square area;
the check-in positions in the check-in record set of the user correspond to the grid cells to which the user belongs one by one, all interest points in each grid cell are represented by the grid cell at the moment, and the check-in positions of the user are presented in a square grid form;
step 2, calculating the influence implicit expression of the social friends;
sequentially inputting the user sign-in record set into a recurrent neural network model according to a time sequence;
at the time t, calculating a geographic coefficient according to the geographic distance between the user and the friend, and performing weighted aggregation on the geographic coefficient and the hidden state of the friend by using a linear function to obtain linear aggregation of the hidden state of the social friend of the user at the time t;
then activating the linear aggregation by utilizing a nonlinear function to obtain an influence implicit expression of the social friends at the moment;
step 3, dynamically updating the implicit access preference of the user;
dynamically updating the implicit access preference of the user by utilizing a hidden layer of the recurrent neural network at the moment of t + 1;
the updating process comprises the following steps: fusing the implicit access preference of the user at the time t, the implicit expression of the influence of social friends at the time t and the position vector of the user at the time t +1, and calculating the implicit access preference of the user at the time t +1 by using a recurrent neural network hidden layer;
step 4, carrying out model training on the recurrent neural network model;
considering the user access position prediction problem as a classification problem, and solving the probability distribution of the next access position of the user;
on the basis, an objective function of the recurrent neural network model is constructed, and model parameters of the recurrent neural network model are optimized by minimizing the objective function, so that the trained recurrent neural network model is obtained;
step 5, predicting the next access position of the user;
on the basis of the step 4, calculating the access probability of the user aiming at each candidate position by utilizing a trained recurrent neural network model for predicting the next access position of the user, and sequencing the access probabilities of all the candidate positions in a descending order;
and finally, selecting candidate positions corresponding to the top-K probability value to form an ordered list as a prediction result.
2. The method of predicting user visiting locations in a geo-sensitive dynamic social environment of claim 1,
the method is characterized in that the step 2 specifically comprises the following steps:
defining t time user uiU with friendsjHas a geographical distance of li,jThe geographic coefficient r is calculated as followst i,j；
Figure FDA0003146530810000011
Wherein a represents a smoothing factor;
defining t time user uiFriend u ofjHas an implicit state of hj tUser uiThe hidden states of the friends are linearly aggregated according to the geographic coefficient to obtain the user uiLinear aggregation of hidden states of good friends at time t Hi t；
Figure FDA0003146530810000021
Wherein Ni is user uiA set of social friends of;
then, for Hi tCarrying out nonlinear activation to obtain the influence implicit expression a of the social friendsi t，
Figure FDA0003146530810000022
Wherein the function
Figure FDA0003146530810000023
Representing a non-linear activation function.
3. The method of predicting user visiting locations in a geo-sensitive dynamic social environment of claim 2,
the method is characterized in that the step 3 specifically comprises the following steps:
at time t +1, the input variables of the recurrent neural network include not only the user uiPosition vector e at time t +1i vt+1And the method also comprises an influence implicit expression a of the social friends at the moment ti tAnd implicit Access preferences h of the user at time ti t；
Defining a user u at time t +1iHas an implicit access preference of hi t+1Then h isi t+1The update method of (1) is as follows:
Figure FDA0003146530810000024
wherein, the function GRU (-) represents a calculation unit of the gated recurrent neural network, and ^ represents the splicing operation.
4. The method of claim 3 for predicting user visiting locations in a geo-sensitive dynamic social environment,
the method is characterized in that the step 4 specifically comprises the following steps:
defining user uiSet of check-in records LiThe access position subsequence from the middle to the t-th moment is Li tAnd obtaining the probability distribution of the next visit position of the user by utilizing a softmax function, wherein the calculation process is as follows:
P(vi t+1＝vm|Li t)＝softmax(WT m·hi t)；
wherein v ismDenotes the m-th candidate position, vmE.g. V, V represents a position set;
vi t+1represents the user u at the time t +1iThe access location of (2);
P(vi t+1＝vm|Li t) The physical meaning of (A) is:
given user u before time tiAccess position sub-sequence L ofi tUser uiThe access position is v at time t +1mThe probability of (d);
w denotes a weight matrix for classification, WmRepresents the mth row of the weight matrix W;
constructing an objective function, wherein a negative log-likelihood Loss function Loss of all user check-in sets is expressed as:
Figure FDA0003146530810000025
wherein, P (v)i n+1|Li n) The physical meaning of (A) is: given user uiIs a subsequence L of the first n access positions ofi nUser uiIs vi n+1The probability of (d);
Lirepresenting user uiU represents a mobile social network user set;
and minimizing the Loss function Loss by using an adaptive moment estimation algorithm, completing autonomous step learning of gradient descent, obtaining model parameters of the recurrent neural network model, and obtaining the trained recurrent neural network model.
5. The method of predicting user visiting locations in a geo-sensitive dynamic social environment of claim 4,
the method is characterized in that the step 5 specifically comprises the following steps:
step 5.1. for target user uiAnd arbitrary candidate position vo，voE to V, and calculating the target user u by utilizing the recurrent neural network model trained in the step 4iThe next access position is voAccess probability P (v)o|Li)；
Wherein, P (v)o|Li) Representing a given user uiSet of check-in records LiThe next access position of the user is voThe probability of (d);
step 5.2, sequencing the access probabilities of all the candidate positions in a descending order;
and 5.3, selecting candidate positions corresponding to the top-K probability value to form an ordered list as a final prediction result.
CN202110578856.7A 2021-05-26 2021-05-26 Prediction method for user access position in geographic sensitive dynamic social environment Active CN113010803B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110578856.7A CN113010803B (en) 2021-05-26 2021-05-26 Prediction method for user access position in geographic sensitive dynamic social environment

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110578856.7A CN113010803B (en) 2021-05-26 2021-05-26 Prediction method for user access position in geographic sensitive dynamic social environment

Publications (2)

Publication Number Publication Date
CN113010803A CN113010803A (en) 2021-06-22
CN113010803B true CN113010803B (en) 2021-08-24

Family

ID=76380838

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110578856.7A Active CN113010803B (en) 2021-05-26 2021-05-26 Prediction method for user access position in geographic sensitive dynamic social environment

Country Status (1)

Country Link
CN (1) CN113010803B (en)

Citations (3)

* Cited by examiner, † Cited by third party
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
CN111723306A (en) * 2020-05-29 2020-09-29 中国科学技术大学智慧城市研究院(芜湖) Sequence position recommendation method based on geographical perception
CN112784163A (en) * 2021-01-29 2021-05-11 东北财经大学 RS 24-based personalized POIs recommendation method

Patent Citations (3)

* Cited by examiner, † Cited by third party
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
CN111723306A (en) * 2020-05-29 2020-09-29 中国科学技术大学智慧城市研究院(芜湖) Sequence position recommendation method based on geographical perception
CN112784163A (en) * 2021-01-29 2021-05-11 东北财经大学 RS 24-based personalized POIs recommendation method

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
基于时空数据挖掘的位置预测与服务推荐研究;胡毅通;《中国优秀硕士学位论文全文数据库 信息科技辑》;20170315(第3期);第16-42页 *

Also Published As

Publication number Publication date
CN113010803A (en) 2021-06-22

Similar Documents

Publication Publication Date Title
CN110928993B (en) User position prediction method and system based on deep cyclic neural network
He et al. Graph attention spatial-temporal network with collaborative global-local learning for citywide mobile traffic prediction
CN113111581B (en) LSTM trajectory prediction method combining space-time factors and based on graph neural network
CN105532030B (en) For analyzing the devices, systems, and methods of the movement of target entity
Jiao et al. A novel next new point-of-interest recommendation system based on simulated user travel decision-making process
CN109948066B (en) Interest point recommendation method based on heterogeneous information network
CN110334293B (en) Position social network-oriented position recommendation method with time perception based on fuzzy clustering
CN112380426A (en) Interest point recommendation method and system based on graph embedding and user long-term and short-term interest fusion
CN112541302A (en) Air quality prediction model training method, air quality prediction method and device
Du et al. A geographical location prediction method based on continuous time series Markov model
Li et al. A collaborative filtering recommendation method based on discrete quantum-inspired shuffled frog leaping algorithms in social networks
Rahimi et al. Behavior-based location recommendation on location-based social networks
Zhao et al. Incorporating spatio-temporal smoothness for air quality inference
Solomon et al. Analyzing movement predictability using human attributes and behavioral patterns
Liu et al. Dynamic metric embedding model for point-of-interest prediction
Xu et al. Deep convolutional recurrent model for region recommendation with spatial and temporal contexts
Quan et al. An optimized task assignment framework based on crowdsourcing knowledge graph and prediction
CN117271899A (en) Interest point recommendation method based on space-time perception
CN103838964A (en) Social relationship network generation method and device based on artificial transportation system
CN113010803B (en) Prediction method for user access position in geographic sensitive dynamic social environment
CN116150511A (en) Next position recommending method based on space-time context and category preference
Zhou et al. Traffic conduction analysis model with time series rule mining
CN115495661A (en) Self-adaptive interest point recommendation method based on long-term and short-term preference of user
CN115510333A (en) POI prediction method based on space-time perception and combined with local and global preferences
Al-Molegi et al. Regions-of-interest discovering and predicting in smartphone environments

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