CN110826698A - Method for embedding and representing crowd moving mode through context-dependent graph - Google Patents

Method for embedding and representing crowd moving mode through context-dependent graph Download PDF

Info

Publication number
CN110826698A
CN110826698A CN201911066089.0A CN201911066089A CN110826698A CN 110826698 A CN110826698 A CN 110826698A CN 201911066089 A CN201911066089 A CN 201911066089A CN 110826698 A CN110826698 A CN 110826698A
Authority
CN
China
Prior art keywords
check
point
track
classifier
training
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.)
Pending
Application number
CN201911066089.0A
Other languages
Chinese (zh)
Inventor
刘芳
钟婷
周帆
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.)
University of Electronic Science and Technology of China
Original Assignee
University of Electronic Science and Technology of China
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 University of Electronic Science and Technology of China filed Critical University of Electronic Science and Technology of China
Priority to CN201911066089.0A priority Critical patent/CN110826698A/en
Publication of CN110826698A publication Critical patent/CN110826698A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • 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
    • G06N3/045—Combinations of networks
    • 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
    • G06F16/90—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
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • 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
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • 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 provides a novel method for representing the movement of people through a context-dependent graph embedding model. The method comprises the steps of firstly generating a fully-connected context correlation graph according to historical tracks of a user, firstly realizing the application of a word embedding technology to a graph model, learning to fuse the sign-in point vector representation of context semantic information, further obtaining the vector representation of each track through a recurrent neural network, and then introducing a reinforcement learning method to find out a generator of the current track. In the whole training process of the model, the method synchronously trains by using the tracks with the labels and the tracks without the labels, and meanwhile, the method also updates the model parameters by using a strategy gradient method in the counterstudy, thereby further enhancing the performance of the model.

Description

Method for embedding and representing crowd moving mode through context-dependent graph
Technical Field
The invention belongs to the field of Neural networks in machine learning, and relates to a deep learning-based method, which is characterized in that a word embedding technology is applied to a graph model for the first time, the sign-in point vector representation of context semantic information is learnt to be fused, the vector representation of each track is obtained through a Recurrent Neural Network (RNN), and then the implicit movement mode of a user is extracted from track data with a label and without the label by utilizing reinforcement learning, so that the user generating the current track is found out.
Background
In a Social networking service (lbs n) application Based on geographic Location information, such as personalized Location recommendation, travel planning, and crowd movement prediction, learning a crowd movement law is an important and complex task, and the task has a very wide application scenario, for example: and finding out the user generating the current track according to the learned crowd mobility rule, and further finding out the friendship of the user, thereby realizing the prediction of the corresponding friendship of the current track. In terms of practical application, the invention can be used to help find extreme organizations in society, such as: given a set of anonymous tracks, identifying a set of criminals/terrorists, or their partnerships, using only the track information; it can also be used for friendship reconstruction, i.e. reconstructing social circles from the respective trajectories.
In recent years, deeply learned models such as Recurrent Neural Networks (RNNs) often utilize low-dimensional embedding techniques to model trajectory information, such as embedded representations of learned check-in points or embedded representations of trajectories. RNN-based methods basically capture semantic information from those continuous variable-length sequences of check-in points, and each RNN stores information from previous units that affects subsequent units. Gradient descent algorithms, such as SGD, Adam, and AMSGRAD, are then typically used to optimize the learnable parameters for a particular task in the RNN model.
However, existing work still faces the following challenges in effectively modeling user movement patterns: (1) there is a lack of a corpus for learning the check-in point embedded representation. The difference between spatio-temporal data processing and word embedding in the natural language processing field is that a rich corpus can be used to train dense representations of different words in the text processing process of natural language, and semantic information contained in each word representation is different for different natural language processing tasks. However, for social network data based on geographic location, semantic information contained in the social network data is obviously insufficient, because most users usually only access a small number of interest points, and because of irregularity of movement tracks and sparseness of check-in points, it is difficult to represent a certain check-in point by using the existing interest points. (2) The restrictions of the context information are ignored. In a given trajectory, different check-in points have their respective geographic location information and check-in time information, which means that each check-in point is limited by the context space affected by the user's movement patterns or check-in preferences, such as: the user's next check-in point may be some point around the current check-in point, and the user prefers to access points of interest that are closer to himself. As can be seen, geographic constraints play an important role in the embedded representation of check-in points and the embedded representation of tracks. (3) Labeled training data is lacking. In many real application scenarios, it is difficult to acquire a large amount of labeled data training models, but this is also the key in the learning of user movement patterns.
Based on the problems, the invention provides a new method for representing crowd movement through context-dependent graph embedding, and the method introduces a reinforced learning idea in the learning process, fully utilizes the geographical position information of the check-in point and the track information without a label, thereby effectively relieving the problems of data sparseness, limited tagged data and the like.
Disclosure of Invention
The invention aims to learn the expression of the movement pattern of a crowd through a new context-dependent graph embedding model, and further introduce a reinforcement learning method to find out a generator of a current track. In the whole training process of the model, the method synchronously trains the tracks with the labels and the tracks without the labels, and simultaneously updates the model parameters by a strategy gradient method in the counterstudy, thereby further enhancing the performance of the model.
Based on the above inventive concept, the present invention designs a method for representing crowd movement patterns by context-dependent graph embedding, which comprises the following steps:
s1, preprocessing data: cutting the original track according to a specific time interval and dividing a training set DLAnd test set DU；
S2, synthesizing a context correlation graph: according to DLAnd DURespectively constructing check-in point space diagrams G by historical tracks of usersspatialAnd access sequence diagram GvisitThen, a fully connected context correlation graph G is synthesized by the two graphstotal；
S3, for the model defined in S31, S32, S33, perform back propagation until the model converges:
s31, context dependent graph embedding: for DLAnd DUIn each check-in point lτBased on the context correlation graph GtotalGenerating a low-dimensional distributed representation C (l) of check-in pointsτ) Then, introducing nonlinear transformation to cooperate with the distribution representation of the training check-in point, and further obtaining the embedded representation C' (l) of the initialized check-in pointτ). Then, the invention makes the time vector u (l) corresponding to the check-in pointτ) Embedded representation C' (l) with initialization check-in pointτ) Splicing to obtain the final representation C of the check-in pointτ；
S32, track coding: based on the final representation of the check-in point obtained in the previous step, the track can be coded by using a recurrent neural network, and the current track is represented by a result X after the last hidden state normalization of the RNN;
s33, pre-training trajectory classifier, discriminator: based on the track embedding representation obtained in the previous step, the marked track (positive sample) { (T) is utilizedl,Ul)∈DLPre-training trajectory classifier pθWherein T islBeing sub-tracks of users, UlLabels for track correspondences, DLRepresenting a set of positive samples. Next, the classifier is used to classify the unlabeled trajectory (negative examples)
Figure BDA0002259398430000031
Predict its corresponding label
Figure BDA0002259398430000032
The discriminator F is then pre-trained by minimizing the cross entropy of the classifier on positive and negative samplesε；
S4, updating the model parameters by using a small batch training method, and repeating the steps until the model traverses all data: from positive samples DLIn the collection of B group data pairs (labeled tracks, corresponding labels), from the negative examples DUData of group B (unlabeled trajectories) were also acquired and passed through a classifier pθPredicting the corresponding labels of the unmarked tracks; using a discriminator FεJudging the credibility of the label predicted by the classifier, and updating the behavior value function a (-) of the discriminator along with the credibility; calculating a gradient of the desired reward
Figure BDA0002259398430000036
Updating parameters { theta, omega, phi } in the model through the strategy gradient, wherein theta is a parameter of the track classifier, omega is a parameter for embedding the sign-in point, and phi is a parameter for embedding the track;
s5, giving an unmarked trackUsing the above-described trained classifier pθAnd discriminator FεPredicting the probability distribution vector corresponding to the track
Figure BDA0002259398430000034
Figure BDA0002259398430000035
Each dimension represents the probability value that the user corresponding to the dimension is the prediction result of the time, and the user corresponding to the dimension with the maximum probability value is selected as the final prediction result.
In the above method for representing the movement pattern of the crowd through the context-dependent graph embedding, the step S1 aims to cut the track and divide the training set DLAnd test set DU. Utensil for cleaning buttockThe method comprises the following steps:
and S11, dividing a complete track into continuous sub-tracks according to a fixed time interval (6 hours), wherein each sub-track is a check-in point sequence divided based on the time interval. If the current trajectory is known by user ukIf it is generated, the trace is recorded asWhere N represents the number of users, each lτ(τ∈[1,n]) Representative time stamp tτA point of attendance, and each lτAll have geographic coordinates corresponding thereto
Figure BDA0002259398430000042
Figure BDA0002259398430000043
Representing the warp and weft values, respectively. For tracks that do not know who produced, i.e. unmarked tracks, they are noted as
Figure BDA0002259398430000044
S12, in order to further mine the moving mode of the user, the invention adds a time factor. Specifically, one day is divided into 24 discrete time stamps and the 24 density-based time representations are sampled in a Gaussian distribution, i.e., at tτNodes that are time-checked-in and their corresponding time vectors may be represented as
Figure BDA0002259398430000045
Wherein d is*A dimension representing a time vector;
s13, dividing training set and testing set, selecting 50% T of each user sub-tracklAnd its corresponding tag UlAs a training set { (T)l,Ul)∈DLThe remaining 50% was used as the test set
Figure BDA0002259398430000046
S14, obtaining a vector representation of the known label: for a known label of the training set data, the present invention will use 0-1 vector encoding to represent the label;
s15, sorting the traces according to their lengths so that the trace lengths of the same batch are as similar as possible. This operation may reduce unnecessary padding during processing of the trace.
In the above method for representing the movement pattern of people by embedding the context-dependent graph, the step S2 is to synthesize a fully connected context-dependent graph Gtotal. The method specifically comprises the following steps:
s21, constructing a check-in point space map, and determining whether two points in the check-in point set should be connected according to whether the distance between the two points is greater than a given threshold value, i.e. constructing a non-directional space map Gspatial{V,EsWhere V is the set of check-in points sampled from the check-in records of the training set and test set, EsRepresenting a collection of edges between nodes, but if a given threshold is too small, some distant check-in points may become outliers, causing the graph to not communicate. In order to avoid the appearance of isolated points, the invention introduces three virtual nodes, namely a starting point sp, a terminal point ep and a completion node cp, and further connects all the nodes in V with the three nodes one by one;
s22, constructing an access sequence diagram, and further constructing a directed diagram G in order to integrate the access intention and the serialized moving mode of the uservisit{V,A,EvisitIn which EvisitRepresenting conversion of places visited by the user, A being GvisitSet of neutron trajectories, for example: if it is
Figure BDA0002259398430000051
Then any check-in point v it containsi∈V；
S23, synthesizing a fully connected context correlation graph Gtotal{ V, E }, first with GspatialInitialization GtotalThen E isvisitThe edges in the graph E are added into the graph E one by one, and an undirected graph G for learning the sign-in point representation is obtainedtotalThe map takes into account both the check-in point geographical location information and the sequence pattern of the check-in.
In the method for representing the movement pattern of the crowd through the context-dependent graph embedding, the step S31 aims to obtain the final representation C of the check-in point through the context-dependent graph embeddingτ. The method specifically comprises the following steps:
s311, in the context correlation graph GtotalContext Con (l) for getting check-in point by random walkτ)＝lτ-a:lτ+aWherein l isτ-a:lτ+aRepresents a check-in point lτNearby continuous 2a +1 check-in point sequences, wherein a is the size of a fixed sliding window, and tau is the index position of the current check-in point in the track;
s312, obtaining a check-in point l by using a continuous bag-of-words modelτLow dimensional distribution of (a) represents C (l)τ)∈RdD is the dimension of the low-dimensional vector space;
s313, introducing nonlinear transformation to cooperatively train the representation of each check-in point, and further obtaining an embedded representation C' (l) of the initialized check-in pointτ)＝Relu(W'C(lτ) + b'), wherein C (l)τ) For signing in point lτThe vector of (1) represents, Relu is a specific activation function, and ω ═ { W ', b' } is a parameter which can be learned in the process of embedding and representing the check-in point;
s314, embedding the initialization of each check-in point into a representation C' (l)τ) And time information u (l)τ) Final representation C of check-in points obtained by splicingτFormally represented as Cτ＝[C'(lτ),u(lτ)]。
In the method for embedding and representing the crowd movement pattern by the context-dependent graph, the step S32 aims to encode the trajectory by using RNN, and further obtain the final representation X of the current trajectory. The method specifically comprises the following steps:
s321, for each (sub) track T ═ l1,l2,...,lnIs given byτ-1And hτ(τ e {1, 2.. eta., n }) represents the previous step and the current hidden state of the trace, respectively, with an initial oneState h0Is a vector of all 0 s. The input gate, forgetting gate and output gate of RNN can be obtained by the following calculation: i.e. iτ＝σ(WiCτ+Uihτ-1+Vicτ-1+bi)，fτ＝σ(WfCτ+Ufhτ-1+Vfcτ-1+bf)，oτ＝σ(WoCτ+Uohτ-1+Vocτ-1+bo) Wherein b is*Is a deviation vector, σ is a logistic activation function, CτRepresents a check-in point lτIs finally expressed, matrix W*，U*，V*Are different door parameters. Memory cell cτThe update rule of (1) is: c. Cτ＝fτcτ-1+iτtanh(WcCτ+Uchτ-1+bc) Where tanh (. cndot.) is a hyperbolic tangent function, initial state c0Is also an all-0 vector, hidden state htThe update rule is as follows: h isτ＝tanh(cτ)oτ；
S322, the invention selects the result after the last hidden state normalization of the current track as the final representation of the track:
Figure BDA0002259398430000061
wherein
Figure BDA0002259398430000062
For the last hidden state hnD' is a hidden state hnOf the dimension, [,]representing a stitching operation between vector elements.
In the above method for representing the movement pattern of the crowd through the context-dependent graph embedding, the step S33 is to utilize the tagging track (positive sample) { (T)l,Ul)∈DLTraining trajectory classifier pθAnd further using the classifier as an unlabeled trajectory (negative example)
Figure BDA0002259398430000063
Predict its corresponding label
Figure BDA0002259398430000064
The discriminator F is then pre-trained by minimizing the cross entropy of the classifier on positive and negative samplesε. The method specifically comprises the following steps:
s331, using the marked track (positive sample) { (T)l,Ul)∈DLPre-training trajectory classifier pθWith training objectives to maximize the desired reward Rθ＝∑u∈UΡθ(u | X) a (X, u), wherein, pθ(u | X) represents the probability that the predicted tag for track X is u, and a (-) is a behavioral cost function for selecting u as the current predicted tag, which is used to measure or evaluate the authenticity of the predicted tag, where a (X, u) ═ 1 if X is data sampled from the labeled track, and otherwise, a (X, u) ═ Fε(X, u), order
Figure BDA0002259398430000065
Reward-based impact classifier as an unmarked trace, W "and b" are learning parameters, o ═ Relu (X)TWru), Relu is the activation function, WrIs a learned parameter matrix;
s332, using classifier pθAs an unmarked track (negative sample)
Figure BDA0002259398430000066
Predict its corresponding label
Figure BDA0002259398430000067
S333, pre-training the discriminator F by minimizing the cross entropy of the classifier on positive and negative samplesεI.e. by
Figure BDA0002259398430000071
In the above method for representing the crowd moving pattern by embedding the context-dependent graph, the step S4 is to continuously update the model parameters { θ, ω, Φ } by using a small-batch training method, and repeat the step until the model traverses all data. The method specifically comprises the following steps:
s41, dividing the small batch data: from training data { (T)l,Ul)∈DLB groups of real labeling tracks and labels corresponding to the real labeling tracks are collected, and test data are obtainedB sets of unlabeled trajectory data and the labels predicted for it by the model are also collected
Figure BDA0002259398430000073
B is the sample size of each small batch;
s42, updating the behavior merit function of the discriminator, i.e. passing through the discriminator Fε(X, u) compute the classifier predicted label Ppθ(u | X) is the probability of a true tag;
s43, calculating a gradient of the desired reward: let the vector of the policy parameter be θ, the update of the policy parameter can be expressed as:
Figure BDA0002259398430000074
wherein the content of the first and second substances,
Figure BDA0002259398430000075
i.e. U is a real tag UlAnd predicted labels
Figure BDA0002259398430000076
The training target of the classifier can be rewritten asIn the training process, the invention uses a small batch training method to update the parameter theta, and the approximate gradient of the parameter theta can be obtained by the following formula:
Figure BDA0002259398430000078
wherein B is the batch size, ul，Xl，
Figure BDA0002259398430000079
Respectively, a true tag, a tagged track, a predicted tag, an untagged track. Therefore, the update rule of the parameter θ is:
Figure BDA00022593984300000710
where β is the learning rate, θnewAnd thetaoldRespectively representing the updated parameter and the current parameter;
s44, in the present invention, in order to better encode the check-in point and the track, the parameter ω embedded to represent the check-in point and the parameter φ embedded to represent the track are updated simultaneously, so the parameter updating formula can be rewritten as
Figure BDA00022593984300000711
Where γ is { θ, ω, Φ }.
Compared with the prior art, the invention has the following beneficial effects:
1. the invention tries to learn the embedded representation of the user check-in point by using a model based on a graph for the first time, and the method provides a new visual angle for a task of learning a mobile mode by using geographical position information and unmarked mobile data;
2. the invention provides a method which integrates context semantic information and can dynamically update the sign-in point embedded representation and the track embedded representation, the method not only provides more accurate and more comprehensive mobile mode representation, but also effectively relieves the problems of data sparseness and limited tagged data;
3. the invention provides a method for extracting an implicit moving mode from tagged and untagged track data by using reinforcement learning, which can approximate real user track distribution, designs a discriminator by using the thought of counterstudy to judge the authenticity of the generated data and further enhances the model performance.
Drawings
FIG. 1 is an overall framework diagram of a method of representing crowd movement patterns through context-dependent graph embedding.
Interpretation of terms
LBSN is an abbreviation for Location-based Social Network, representing a "Location-based Social Network". The LBSN not only contains the relationship between people in the traditional social network, but also records the information of the user sign-in time, the geographic position and the like.
RNN is an abbreviation for current Neural Network, denoted "Recurrent Neural Network", a class of Neural networks used for processing sequence data.
Detailed Description
The invention is further described below with reference to the accompanying drawings.
Examples
The method for representing crowd movement patterns through context-dependent graph embedding provided by the present implementation can be used for such data sets containing user check-in records. Real world LBSN data set as shown in Table 1, in Gowalla: (B)http:// snap.stanford.edu/data/loc-gowalla.htmlAcquisition) was performed as an example.
Table 1: information related to the experimental data of the invention
Figure BDA0002259398430000081
| u | represents the number of users;
|DL|/|DUl respectively represents the number of sub-tracks used for training and testing;
l represents the number of different check-in points in the dataset;
Figure BDA0002259398430000091
representing the average length of the track before being divided;
Trrepresenting a range of track lengths in the data set;
new York represents data in the data set Foursquare with a New York area;
tokyo represents data in the data set Foursquare with an area of Tokyo.
In this embodiment, the present invention represents people by context-dependent graph embedding, as shown in FIG. 1The method of the group moving mode includes steps S1 to S5. Step S1, aiming at the Gowalla data set, firstly dividing the track to obtain a (sub) track set and a check-in point set, then connecting each sub track related by each user with the corresponding user, and selecting 50% of the marked tracks of the total sub tracks of each user as a training set DLThe remaining 50% unmarked sub-tracks DUIs used for testing; step S2, respectively constructing check-in point space graph G according to the historical track of the userspatialAnd access sequence diagram GvisitThen, a fully connected context correlation graph G is synthesized by the two graphstotal(ii) a Step S3, for the model defined in S31, S32, S33, performs back propagation until the model converges: s31, for DLAnd DUIn each check-in point lτBased on the context correlation graph GtotalGenerating a low-dimensional distributed representation C (l) of check-in pointsτ) Then, introducing nonlinear transformation to cooperate with the distribution representation of the training check-in point to obtain the embedded representation C' (l) of the initialized check-in pointτ) Then, the present invention provides the time vector u (l) corresponding to the check-in pointτ) Embedded representation C' (l) with initialization check-in pointτ) Splicing to obtain the final representation C of the check-in pointτ(ii) a S32, based on the embedded expression of the check-in point obtained in the previous step, the track can be coded by using a recurrent neural network, wherein the current track is expressed by the result X after the last hidden state normalization of the RNN; s33, based on the trace embedding representation obtained in the previous step, using the marked trace (positive sample) { (T)l,Ul)∈DLPretraining a trajectory classifier, initializing the trajectory classifier pθAnd further using the classifier as an unlabeled trajectory (negative example)Predict its corresponding label
Figure BDA0002259398430000093
The discriminator F is then pre-trained by minimizing the cross entropy of the classifier on positive and negative samplesε(ii) a Step S4, updating the model parameters using a small batch training method, and repeating the step until the model traverses all data: from the tagging track DLIn the collection of B group data pairs (labeled tracks, corresponding labels), from unlabeled track DUData of group B (unlabeled trajectories) were also acquired and passed through a classifier pθPredicting the corresponding labels of the unmarked tracks; using a discriminator FεJudging the credibility of the label predicted by the classifier, and updating the behavior value function a (-) of the discriminator along with the credibility; calculating a gradient of the desired reward
Figure BDA0002259398430000101
Updating parameters { theta, omega, phi } in the model through the strategy gradient; step S5, an unmarked track is given
Figure BDA0002259398430000102
Using the above-described trained classifier pθAnd discriminator FεPredicting the probability distribution vector corresponding to the track
Figure BDA0002259398430000103
Each dimension represents the probability value that the user corresponding to the dimension is the prediction result of the time, and the user corresponding to the dimension with the maximum probability value is selected as the final prediction result.
As can be seen from the experimental results in table 2, the prediction effect of the method for representing the movement pattern of the crowd through the context-dependent graph embedding of the present invention is completely superior to that of the conventional method. Therefore, the graph-based method which is provided by the invention and integrates the context semantic information and can dynamically update the check-in point embedded representation and the track embedded representation actually improves the prediction performance of the model.
ACC @ K represents the accuracy, namely the proportion of the number of predicted correct tracks to the number of all tracks;
macro-P represents the accuracy rate, namely the proportion of the number of tracks which are predicted correctly to the number of tracks which are predicted;
macro-R represents the recall rate, namely the proportion of the number of predicted correct tracks to the number of real tracks;
macro-F1 represents the harmonic mean of the exact value and recall;
TULER-LSTM, TULER-GRU, Bi-TULER: the present invention replicates three TULER methods based on LSTM, GRU, and bi-directional LSTM units, respectively, TULER being the best method to perform on the user-trajectory link problem in the near future.
HTULER-L, HTULER-G, HTULER-B: three htule methods, based on multi-layer LSTM, GRU, and bi-directional LSTM cells, respectively, are reproduced for purposes of the present invention.
TULVAE is a method to solve the user-trajectory linkage problem in a semi-supervised manner by variational inference. TLUTE-L, TLUTE-G: TLUTE was originally a method to measure the similarity between a user's embedded representation and the embedded representation of a track, but the task of the present invention is to link a track to the corresponding user, so the task can be considered as a classification task. Here, for the convenience of comparison with the method proposed by the present invention, the present invention removes the module of similarity measure in the original TLUTE, and TLUTE-L and TLUTE-G are simplified version TLUTE methods based on LSTM and GRU units, respectively.
Table 2: effects of GraphTUL on track-user Link prediction on four real datasets
Figure BDA0002259398430000111
It will be appreciated by those of ordinary skill in the art that the embodiments described herein are intended to assist the reader in understanding the principles of the invention and are to be construed as being without limitation to such specifically recited embodiments and examples. Those skilled in the art can make various other specific changes and combinations based on the teachings of the present invention without departing from the spirit of the invention, and these changes and combinations are within the scope of the invention.

Claims (7)

1. A method for representing movement patterns of a population by context-dependent graph embedding, comprising the steps of:
s1, preprocessing data: cutting the original track according to a specific time interval and dividing a training set DLAnd test set DU；
S2, synthesizing a context correlation graph: according to DLAnd DURespectively constructing check-in point space diagrams G by historical tracks of usersspatialAnd access sequence diagram GvisitThen, a fully connected context correlation graph G is synthesized by the two graphstotal；
S3, for the model defined in S31, S32, S33, perform back propagation until the model converges:
s31, context dependent graph embedding: for DLAnd DUIn each check-in point lτBased on the context correlation graph GtotalGenerating a low-dimensional distributed representation C (l) of check-in pointsτ) Then, introducing nonlinear transformation to cooperate with the distribution representation of the training check-in point, and further obtaining the embedded representation C' (l) of the initialized check-in pointτ). Then, the invention makes the time vector u (l) corresponding to the check-in pointτ) Embedded representation C' (l) with initialization check-in pointτ) Splicing to obtain the final representation C of the check-in pointτ；
S32, track coding: and coding the track by utilizing a recurrent neural network based on the final representation of the check-in point obtained in the previous step. In this case, the present invention uses the result X after the last hidden state normalization of RNN to represent the current trace;
s33, pre-training trajectory classifier, discriminator: based on the trace-embedding representation obtained in the previous step, using the scaled trace, i.e., { (T) positive samplesl,Ul)∈DLPre-training trajectory classifier pθWherein T islBeing sub-tracks of users, UlLabels for track correspondences, DLRepresenting a set of positive samples. Subsequently, the classifier is used to classify the unlabeled trajectory, i.e. the negative examples
Figure FDA0002259398420000011
Predict its corresponding label
Figure FDA0002259398420000012
The discriminator F is then pre-trained by minimizing the cross entropy of the classifier on positive and negative samplesε；
S4, updating the model parameters by using a small batch training method, and repeating the steps until the model traverses all data: from positive samples DLIn the collection of group B data pairs, from negative samples DUB group data are also acquired and passed through a classifier pθPredicting the corresponding labels of the unmarked tracks; using a discriminator FεJudging the credibility of the label predicted by the classifier, and updating the behavior value function a (-) of the discriminator along with the credibility; calculating a gradient of the desired reward
Figure FDA0002259398420000021
Updating parameters { theta, omega, phi } in the model through the strategy gradient, wherein theta is a parameter of the track classifier, omega is a parameter for embedding the sign-in point, and phi is a parameter for embedding the track;
s5, giving an unmarked track
Figure FDA0002259398420000022
Using the above-described trained classifier pθAnd discriminator FεPredicting the probability distribution vector corresponding to the track
Figure FDA0002259398420000024
Each dimension represents the probability value that the user corresponding to the dimension is the prediction result of the time, and the user corresponding to the dimension with the maximum probability value is selected as the final prediction result.
2. The method of claim 1, wherein the step S1 comprises the following sub-steps:
s11, dividing a complete track into continuous sub-tracks according to a fixed time interval, such as 6 hours, where each sub-track is a sequence of check-in points divided based on the time interval. If the current trajectory is known by user ukIf it is generated, the trace is recorded as
Figure FDA0002259398420000025
Where N represents the number of users, each lτ(τ∈[1,n]) Representative time stamp tτA point of attendance, and each lτAll have geographic coordinates corresponding thereto
Figure FDA0002259398420000026
Figure FDA0002259398420000027
Representing the warp and weft values, respectively. For tracks that do not know who produced, i.e. unmarked tracks, they are noted as
Figure FDA0002259398420000028
S12, in order to further mine the moving mode of the user, the invention adds a time factor. Specifically, one day is divided into 24 discrete time stamps and the 24 density-based time representations are sampled in a Gaussian distribution, i.e., at tτNodes that are time-checked-in and their corresponding time vectors may be represented asWherein d is*A dimension representing a time vector;
s13, dividing training set and testing set, selecting 50% T of each user sub-tracklAnd its corresponding tag UlAs a training set { (T)l,Ul)∈DLThe remaining 50% was used as the test set
Figure FDA00022593984200000210
S14, obtaining a vector representation of the known label: for a known label of the training set data, the present invention will use 0-1 vector encoding to represent the label;
s15, sorting the traces according to their lengths so that the trace lengths of the same batch are as similar as possible. This operation may reduce unnecessary padding during processing of the trace.
3. The method of claim 1, wherein the step S2 comprises the following sub-steps:
s21, constructing a check-in point space map, and determining whether two points in the check-in point set should be connected according to whether the distance between the two points is greater than a given threshold value, i.e. constructing a non-directional space map Gspatial{V,EsWhere V is the set of check-in points sampled from the check-in records of the training set and test set, EsRepresenting a collection of edges between nodes, but if a given threshold is too small, some distant check-in points may become outliers, causing the graph to not communicate. In order to avoid the appearance of isolated points, the invention introduces three virtual nodes, namely a starting point sp, a terminal point ep and a completion node cp, and further connects all the nodes in V with the three nodes one by one;
s22, constructing an access sequence diagram, and further constructing a directed diagram G in order to integrate the access intention and the serialized moving mode of the uservisit{V,A,EvisitIn which EvisitRepresenting conversion of places visited by the user, A being GvisitSet of neutron trajectories, for example: if it is
Figure FDA0002259398420000031
Then any check-in point v it containsi∈V；
S23, synthesizing a fully connected context correlation graph Gtotal{ V, E }, first with GspatialInitializationGtotalThen E isvisitThe edges in the graph E are added into the graph E one by one, and an undirected graph G for learning the sign-in point representation is obtainedtotalThe map takes into account both the check-in point geographical location information and the sequence pattern of the check-in.
4. The method of claim 1, wherein the step S31 comprises the following sub-steps:
s311, in the context correlation graph GtotalContext Con (l) for getting check-in point by random walkτ)＝lτ-a:lτ+aWherein l isτ-a:lτ+aRepresents a check-in point lτNearby continuous 2a +1 check-in point sequences, wherein a is the size of a fixed sliding window, and tau is the index position of the current check-in point in the track;
s312, obtaining a check-in point l by using a continuous bag-of-words modelτLow dimensional distribution of (a) represents C (l)τ)∈RdD is the dimension of the low-dimensional vector space;
s313, introducing nonlinear transformation to cooperatively train the representation of each check-in point, and further obtaining an embedded representation C' (l) of the initialized check-in pointτ)＝Relu(W'C(lτ) + b'), wherein C (l)τ) For signing in point lτThe vector of (1) represents, Relu is a specific activation function, and ω ═ { W ', b' } is a parameter which can be learned in the process of embedding and representing the check-in point;
s314, embedding the initialization of each check-in point into a representation C' (l)τ) And time information u (l)τ) Final representation C of check-in points obtained by splicingτFormally represented as Cτ＝[C'(lτ),u(lτ)]。
5. The method of claim 1, wherein the step S32 comprises the following sub-steps:
s321, for each sub-track T ═ { l ═ l1,l2,...,lnIs given byτ-1And hτ(τ e {1, 2.. eta., n }) represents the previous step and the current hidden state of the trace, respectively, with an initial state h0Is a vector of all 0 s. The input gate, forgetting gate and output gate of RNN can be obtained by the following calculation: i.e. iτ＝σ(WiCτ+Uihτ-1+Vicτ-1+bi)，fτ＝σ(WfCτ+Ufhτ-1+Vfcτ-1+bf)，oτ＝σ(WoCτ+Uohτ-1+Vocτ-1+bo) Wherein b is*Is a deviation vector, σ is a logistic activation function, CτRepresents a check-in point lτIs finally expressed, matrix W*，U*，V*Are different door parameters. Memory cell cτThe update rule of (1) is: c. Cτ＝fτcτ-1+iτtanh(WcCτ+Uchτ-1+bc) Where tanh (. cndot.) is a hyperbolic tangent function, initial state c0Is also an all-0 vector, hidden state htThe update rule is as follows: h isτ＝tanh(cτ)oτ；
S322, the invention selects the result after the last hidden state normalization of the current track as the final representation of the track:
Figure FDA0002259398420000041
wherein
Figure FDA0002259398420000042
For the last hidden state hnD' is a hidden state hnOf the dimension, [,]representing a stitching operation between vector elements.
6. The method of claim 1, wherein the step S33 comprises the following sub-steps:
s331, utilizing and addingTrace, i.e. positive sample { (T)l,Ul)∈DLPre-training trajectory classifier pθWith training objectives to maximize the desired reward Rθ＝∑u∈UΡθ(u | X) a (X, u), wherein, pθ(u | X) represents the probability that the predicted tag for track X is u, and a (-) is a behavioral cost function for selecting u as the current predicted tag, which is used to measure or evaluate the authenticity of the predicted tag, where a (X, u) ═ 1 if X is data sampled from the labeled track, and otherwise, a (X, u) ═ Fε(X, u), order
Figure FDA0002259398420000051
Reward-based impact classifier as an unmarked trace, W "and b" are learning parameters, o ═ Relu (X)TWru), Relu is the activation function, WrIs a learned parameter matrix;
s332, using classifier pθFor unmarked tracks, i.e. negative examples
Figure FDA0002259398420000052
Predict its corresponding label
Figure FDA0002259398420000053
S333, pre-training the discriminator F by minimizing the cross entropy of the classifier on positive and negative samplesεI.e. by
7. The method of claim 1, wherein the step S4 comprises the following sub-steps:
s41, dividing the small batch data: from training data { (T)l,Ul)∈DLB groups of real labeling tracks and labels corresponding to the real labeling tracks are collected, and test data are obtained
Figure FDA0002259398420000055
B sets of unlabeled trajectory data and the labels predicted for it by the model are also collectedB is the sample size of each small batch;
s42, updating the behavior merit function of the discriminator, i.e. passing through the discriminator Fε(X, u) compute the classifier predicted label Ppθ(u | X) is the probability of a true tag;
s43, calculating a gradient of the desired reward: let the vector of the policy parameter be θ, the update of the policy parameter can be expressed as:
Figure FDA0002259398420000057
wherein the content of the first and second substances,
Figure FDA0002259398420000058
i.e. U is a real tag UlAnd predicted labels
Figure FDA0002259398420000059
The training target of the classifier can be rewritten as
In the training process, the invention uses a small batch training method to update the parameter theta, and the approximate gradient of the parameter theta can be obtained by the following formula:
Figure FDA00022593984200000511
wherein B is the batch size, ul，Xl，
Figure FDA00022593984200000512
Respectively, a true tag, a tagged track, a predicted tag, an untagged track. Therefore, the update rule of the parameter θ is:
Figure FDA00022593984200000513
where β is the learning rate, θnewAnd thetaoldRespectively representing the updated parameter and the current parameter;
s44, in the present invention, in order to better encode the check-in point and the track, the parameter ω embedded to represent the check-in point and the parameter φ embedded to represent the track are updated simultaneously, so the parameter updating formula can be rewritten as
Figure FDA0002259398420000061
Where γ is { θ, ω, Φ }.
CN201911066089.0A 2019-11-04 2019-11-04 Method for embedding and representing crowd moving mode through context-dependent graph Pending CN110826698A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201911066089.0A CN110826698A (en) 2019-11-04 2019-11-04 Method for embedding and representing crowd moving mode through context-dependent graph

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201911066089.0A CN110826698A (en) 2019-11-04 2019-11-04 Method for embedding and representing crowd moving mode through context-dependent graph

Publications (1)

Publication Number Publication Date
CN110826698A true CN110826698A (en) 2020-02-21

Family

ID=69552527

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201911066089.0A Pending CN110826698A (en) 2019-11-04 2019-11-04 Method for embedding and representing crowd moving mode through context-dependent graph

Country Status (1)

Country Link
CN (1) CN110826698A (en)

Cited By (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111783295A (en) * 2020-06-28 2020-10-16 中国人民公安大学 Dynamic identification and prediction evaluation method and system for urban community specific human behavior chain
CN112380426A (en) * 2020-10-23 2021-02-19 南京邮电大学 Interest point recommendation method and system based on graph embedding and user long-term and short-term interest fusion
CN112667763A (en) * 2020-12-29 2021-04-16 电子科技大学 Trajectory prediction method based on self-adaptive timestamp and multi-scale feature extraction
CN113378891A (en) * 2021-05-18 2021-09-10 东北师范大学 Urban area relation visual analysis method based on track distribution representation
CN113378074A (en) * 2021-06-10 2021-09-10 电子科技大学 Social network user trajectory analysis method based on self-supervision learning
CN113705099A (en) * 2021-05-09 2021-11-26 电子科技大学 Social platform rumor detection model construction method and detection method based on contrast learning
CN113781527A (en) * 2021-11-10 2021-12-10 华中科技大学 Pedestrian trajectory prediction method and system based on multi-interaction space-time diagram network
CN114880586A (en) * 2022-06-07 2022-08-09 电子科技大学 Confrontation-based social circle inference method through mobility context awareness
CN117520906A (en) * 2024-01-05 2024-02-06 北京航空航天大学 Crowd classification method and system based on different characteristics of crowd travel activity entropy

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20180121799A1 (en) * 2016-11-03 2018-05-03 Salesforce.Com, Inc. Training a Joint Many-Task Neural Network Model using Successive Regularization
CN109117476A (en) * 2018-07-12 2019-01-01 浙江鸿程计算机系统有限公司 A kind of personalized place method for recognizing semantics based on type multiple-situation insertion
US20190042745A1 (en) * 2017-12-28 2019-02-07 Intel Corporation Deep learning on execution trace data for exploit detection
CN110147892A (en) * 2019-02-20 2019-08-20 电子科技大学 Mankind's Move Mode presumption model, training method and estimation method based on variation track context-aware
CN110232480A (en) * 2019-03-01 2019-09-13 电子科技大学 The item recommendation method and model training method realized using the regularization stream of variation

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20180121799A1 (en) * 2016-11-03 2018-05-03 Salesforce.Com, Inc. Training a Joint Many-Task Neural Network Model using Successive Regularization
US20190042745A1 (en) * 2017-12-28 2019-02-07 Intel Corporation Deep learning on execution trace data for exploit detection
CN110059484A (en) * 2017-12-28 2019-07-26 英特尔公司 Deep learning is executed to be used for Hole Detection to track data
CN109117476A (en) * 2018-07-12 2019-01-01 浙江鸿程计算机系统有限公司 A kind of personalized place method for recognizing semantics based on type multiple-situation insertion
CN110147892A (en) * 2019-02-20 2019-08-20 电子科技大学 Mankind's Move Mode presumption model, training method and estimation method based on variation track context-aware
CN110232480A (en) * 2019-03-01 2019-09-13 电子科技大学 The item recommendation method and model training method realized using the regularization stream of variation

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
FAN ZHOU等: ""DeepLink: A Deep Learning Approach for User"", 《IEEE CONFERENCE ON COMPUTER COMMUNICATIONS》 *
高强等: ""轨迹大数据:数据处理关键技术研究综述"", 《软件学报》 *

Cited By (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111783295A (en) * 2020-06-28 2020-10-16 中国人民公安大学 Dynamic identification and prediction evaluation method and system for urban community specific human behavior chain
CN111783295B (en) * 2020-06-28 2020-12-22 中国人民公安大学 Dynamic identification and prediction evaluation method and system for urban community specific human behavior chain
CN112380426A (en) * 2020-10-23 2021-02-19 南京邮电大学 Interest point recommendation method and system based on graph embedding and user long-term and short-term interest fusion
CN112380426B (en) * 2020-10-23 2023-10-13 南京邮电大学 Interest point recommendation method and system based on fusion of graph embedding and long-term interest of user
CN112667763A (en) * 2020-12-29 2021-04-16 电子科技大学 Trajectory prediction method based on self-adaptive timestamp and multi-scale feature extraction
CN113705099A (en) * 2021-05-09 2021-11-26 电子科技大学 Social platform rumor detection model construction method and detection method based on contrast learning
CN113378891A (en) * 2021-05-18 2021-09-10 东北师范大学 Urban area relation visual analysis method based on track distribution representation
CN113378074A (en) * 2021-06-10 2021-09-10 电子科技大学 Social network user trajectory analysis method based on self-supervision learning
CN113781527A (en) * 2021-11-10 2021-12-10 华中科技大学 Pedestrian trajectory prediction method and system based on multi-interaction space-time diagram network
CN113781527B (en) * 2021-11-10 2022-02-08 华中科技大学 Pedestrian trajectory prediction method and system based on multi-interaction space-time diagram network
CN114880586A (en) * 2022-06-07 2022-08-09 电子科技大学 Confrontation-based social circle inference method through mobility context awareness
CN117520906A (en) * 2024-01-05 2024-02-06 北京航空航天大学 Crowd classification method and system based on different characteristics of crowd travel activity entropy
CN117520906B (en) * 2024-01-05 2024-03-12 北京航空航天大学 Crowd classification method and system based on different characteristics of crowd travel activity entropy

Similar Documents

Publication Publication Date Title
CN110826698A (en) Method for embedding and representing crowd moving mode through context-dependent graph
CN109447140B (en) Image identification and cognition recommendation method based on neural network deep learning
Zhou et al. Personalized learning full-path recommendation model based on LSTM neural networks
Qin et al. A dual-stage attention-based recurrent neural network for time series prediction
Wu et al. Evolving RBF neural networks for rainfall prediction using hybrid particle swarm optimization and genetic algorithm
CN110046671A (en) A kind of file classification method based on capsule network
CN109583565A (en) Forecasting Flood method based on the long memory network in short-term of attention model
CN113068131B (en) Method, device, equipment and storage medium for predicting user movement mode and track
CN106980831A (en) Based on self-encoding encoder from affiliation recognition methods
CN112395502A (en) Page display method and device, electronic equipment and computer storage medium
CN114168819B (en) Post matching method and device based on graph neural network
CN115148302A (en) Compound property prediction method based on graph neural network and multi-task learning
Gao et al. Adversarial mobility learning for human trajectory classification
Gao et al. Contextual spatio-temporal graph representation learning for reinforced human mobility mining
CN115687760A (en) User learning interest label prediction method based on graph neural network
CN115544239A (en) Deep learning model-based layout preference prediction method
Cao et al. A dual attention model based on probabilistically mask for 3D human motion prediction
Gammelli et al. Recurrent flow networks: A recurrent latent variable model for density estimation of urban mobility
Yu et al. Using information entropy and a multi-layer neural network with trajectory data to identify transportation modes
Chen et al. Feature extraction method of 3D art creation based on deep learning
CN114461931A (en) User trajectory prediction method and system based on multi-relation fusion analysis
CN109784404A (en) A kind of the multi-tag classification prototype system and method for fusion tag information
CN113989544A (en) Group discovery method based on deep map convolution network
CN116450704A (en) Automatic generation method and generation device of flow model
Chu et al. Simulating human mobility with a trajectory generation framework based on diffusion model

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
WD01 Invention patent application deemed withdrawn after publication

Application publication date: 20200221

WD01 Invention patent application deemed withdrawn after publication