CN110059144A - A kind of track owner's prediction technique based on convolutional neural networks - Google Patents

Abstract

Track owner's prediction technique based on convolutional neural networks that the invention discloses a kind of, according to the track of all users formed an oriented no weight graph G=<V, E>, pass through Node2Vec learn track position ID low-dimensional real-valued vectors；Then slicing treatment is carried out to user trajectory, position ID is replaced with position ID corresponding low-dimensional real-valued vectors to the random length track after cutting, and form the fixed dimension matrix of track by interception or filling；Then, one four layers of convolutional neural networks of building and training are as prediction model, then by the track Input matrix of user location longitude and latitude to be detected building into trained prediction model, the probability distribution of track owner classification is obtained, the index of maximum value in probability distribution is finally corresponded to the number of owner labeled as the track.

Description

A Trajectory Ownership Prediction Method Based on Convolutional Neural Networks

technical field

The invention belongs to the technical field of machine learning, and more particularly, relates to a trajectory owner prediction method based on a convolutional neural network.

Background technique

Trajectory owner prediction is based on feature extraction and analysis of the trajectory of an unknown owner, and then determines the owner of the trajectory. Trajectory owner prediction is the basis of many location-based services, and is of great significance for improving the quality of location-based services. Service providers can use the prediction results for personalized recommendation and preference-based path planning.

The existing trajectory ownership prediction methods usually treat the trajectory as a time series, and then use RNN and other methods to learn the representation of the time series. Although this method learns the context of trajectories, in a trajectory sequence, a specific position or a combination of certain specific positions may be very important for the classification of trajectories, and existing methods cannot effectively capture these characteristics of trajectories. , and the convolutional neural network can learn these features better, so we propose a trajectory owner prediction method based on the convolutional neural network.

SUMMARY OF THE INVENTION

The purpose of the present invention is to overcome the deficiencies of the prior art and provide a trajectory owner prediction method based on a convolutional neural network, which improves the accuracy of trajectory prediction by improving trajectory modeling and feature extraction methods.

In order to achieve the above purpose of the invention, the present invention proposes a method for predicting trajectory ownership based on a convolutional neural network, which is characterized in that it includes the following steps:

(1), data preprocessing

(1.1) Count the longitude and latitude of the historical location of all user trajectories in chronological order to form a longitude and latitude set, wherein, if a certain longitude and latitude appears repeatedly, it is only reserved once in the longitude and latitude set;

Number each longitude and latitude in the longitude and latitude set from 1, and give a unique location ID;

(1.2), in chronological order, use the location ID given in step (1.1) to replace the longitude and latitude of the historical location of the user track, then each user track is represented by a string of location IDs; at the same time, each user is represented by a unique integer ID Identify, so that the user and the user track can form the owner track in the form of [user ID, track (location ID, ..., location ID)];

(1.3) Form a directed unweighted graph G=<V, E> according to the owner trajectory, where V is the set of all location IDs, if a user appears from ID _i to ID _j , then <ID _i , ID _j > represents a directed edge, and all such edges constitute the edge set E of the graph G;

(2), track representation

(2.1), take the constructed directed unweighted graph G as the input, and learn the low-dimensional real-valued vector of each position ID in G through the Node2Vec algorithm;

(2.2), slice the trajectory of each user in step (1.2) according to fixed time intervals, thereby dividing each user trajectory into several position ID sequences, and then use the user ID to own the cut position ID sequence. identification;

(2.3), to the indefinite length trajectory after cutting, replace the position ID with the low-dimensional real-valued vector corresponding to the position ID, thereby generating the trajectory matrix of each user;

Then, by intercepting or filling, construct a fixed-dimensional trajectory matrix for each user to form a data set, where the filled vector is the average of the low-dimensional real-valued vectors corresponding to all location IDs;

(3), build a prediction model

Construct a four-layer convolutional neural network, whose input layer is a fixed-dimensional trajectory matrix; the convolutional layer sets three convolution kernels m*embedding_size, where m is a constant, and the embedding size is the low-dimensional real-valued vector output by Node2vec. Dimension; the pooling layer is k-max pooling, k represents the first k maximum values after convolution; the output of the fully connected layer is input to the softmax function to obtain the probability distribution of the track owner;

(4), training prediction model

(4.1), build a training set

The fixed-dimensional trajectory matrix set X of some users in the dataset and the corresponding one-hot category vector set Y are used as the training set, where X=[Vec_x ₁ ,Vec_x ₂ ,...,Vec_x _n ], Vec_x _n represents the first Trajectory matrix of n fixed dimensions of users, Y=[Vec_t ₁ , Vec_t ₂ ,...,Vec_t _n ], Vec_t _n represents the one-hot category vector corresponding to the nth owner, if the nth position is 1, Then the rest of the positions are all 0;

(4.2), initialize the prediction model

The value of the weight matrix W _p of each convolution kernel in the initialized convolution layer is a normal distribution with a mean value of 0 and a variance of 0.1; at the same time, the bias vector B _p of each convolution kernel in the convolution layer is initialized to 0.1 , the number of elements of each bias vector is the number of neurons in the corresponding layer; the weight matrix of the initialized fully connected layer is W, and its dimension is [batch_size*k*number of convolution kernels, number of categories], and at the same time initialize The bias vector B value of the fully connected layer is 0.1, and the number of elements is the number of categories; among them, batch_size is a constant, p=1, 2, 3, p represents the number of convolution kernels in the convolution layer;

(4.3) Input the training set into the initialized prediction model, use the Adam algorithm to optimize the loss function, and then use the error backpropagation BP algorithm to pass the error to the previous layer, update the weight matrix W _p , bias of the convolution layer The vector B _p , the weight matrix W of the fully connected layer, and the bias vector B, after several iterations, the converged neural network model is obtained, thereby obtaining the trained prediction model;

(5), trajectory owner prediction

According to the method described in steps (1) and (2), the latitude and longitude of the location of the user to be detected is used to construct a trajectory matrix of the fixed dimension of the user, and then the constructed trajectory matrix is input into the already trained prediction model, and the corresponding trajectory of the trajectory is obtained. The probability distribution of all owner categories, and the index of the maximum value in the probability distribution is the owner number corresponding to the track.

The purpose of the invention of the present invention is achieved in this way:

The present invention is based on the trajectory owner prediction method of the convolutional neural network, forms a directed unweighted graph G=<V, E> according to the trajectory of all users, and learns the low-dimensional real-valued vector of the trajectory position ID through Node2Vec; The trajectory is sliced, and the position ID is replaced by the low-dimensional real-valued vector corresponding to the position ID for the indefinite-length trajectory after cutting, and the fixed-dimensional matrix of the trajectory is formed by intercepting or filling; then, a four-layer convolutional neural network is constructed and trained. As a prediction model, the trajectory matrix constructed by the longitude and latitude of the user's location to be detected is input into the trained prediction model to obtain the probability distribution of the trajectory owner classification, and finally the index of the maximum value in the probability distribution is marked as the corresponding attribute of the trajectory main number.

Meanwhile, the trajectory owner prediction method based on the convolutional neural network of the present invention also has the following beneficial effects:

(1) Construct the owner trajectory sequence into a network, that is, a directed unweighted graph, and learn the low-dimensional real-valued vector of each node in the network through the Node2Vec algorithm;

(2) The vector filled by the trajectory is the average value of all vectors. Compared with all filling of 0, the accuracy of owner prediction has been greatly improved;

(3) Compared with the traditional convolutional neural network, the accuracy of trajectory owner prediction is further improved.

Description of drawings

Fig. 1 is the flow chart of the trajectory owner prediction method based on convolutional neural network of the present invention;

Figure 2 is a schematic diagram of the architecture of the prediction model.

Detailed ways

The specific embodiments of the present invention are described below with reference to the accompanying drawings, so that those skilled in the art can better understand the present invention. It should be noted that, in the following description, when the detailed description of known functions and designs may dilute the main content of the present invention, these descriptions will be omitted here.

Example

FIG. 1 is a flow chart of a method for predicting trajectory ownership based on a convolutional neural network according to the present invention.

In this embodiment, as shown in FIG. 1 , a method for predicting trajectory ownership based on a convolutional neural network of the present invention includes the following steps:

S1, data preprocessing

S1.1. Count the longitude and latitude of the historical locations of all user trajectories in chronological order to form a longitude and latitude set, wherein if a certain longitude and latitude appears repeatedly, it is only reserved once in the longitude and latitude set;

As shown in Table 1, each longitude and latitude in the longitude and latitude set is numbered from 1, and a unique location ID is given;

Table 1 is the location ID identification table of latitude and longitude;

Numbering latitude and longitude 1 39.747652-104.99251 2 39.891383-105.070814 3 39.891077-105.068532 4 39.750469-104.999073 ... ...

Table 1

S1.2. In chronological order, use the location ID given in step S1.1 to replace the latitude and longitude of the historical location of the user track, then each user track is represented by a string of location IDs; at the same time, each user is represented by a unique integer ID Identify, so that the user and the user track can form the owner track in the form of [user ID, track (location ID, ..., location ID)]; in this embodiment, such as a certain owner track: [0, ( 622, 474, 474, 474, 481, 482, 482, 83, 83, 270, 487, 270, 270, 83, 83, 471,...)], where 0 is the user ID, followed by the location ID sequence;

S1.3. Form a directed unweighted graph G=<V, E> according to the owner trajectory, where V is the set of all location IDs, if a user appears from ID _i to ID _j , then <ID _i , ID _j > represents a directed edge, and all such edges constitute the edge set E of the graph G;

S2, track representation

S2.1. Taking the constructed directed unweighted graph G as input, learn the low-dimensional real-valued vector of each position ID in G through the Node2Vec algorithm; Node2Vec is a network representation learning method, and its specific process belongs to the prior art , and will not be repeated here.

In this embodiment, the low-dimensional real-valued vector of the position ID numbered 60 is represented as:

[60(-0.383389, -0.826315, -1.379363, ..., -1.839076, 1.930556, 0.502587)];

S2.2. Slice the trajectory of each user in step S1.2 according to a fixed time interval, so as to divide each user trajectory into several position ID sequences, and then use the user ID to own the cut position ID sequence. In this embodiment, the slicing effect is as follows:

[0, (622, 474, 474, 474, 481, 482, 482, 83)]

[0, (83, 270, 487, 270, 270, 83, 83, 471)]

...

Among them, 0 is the user ID, followed by the location ID sequence;

S2.3, replace the position ID with the low-dimensional real-valued vector corresponding to the position ID for the indefinite-length trajectory after cutting, thereby generating the trajectory matrix of each user;

Then, a fixed-dimensional trajectory matrix of each user is constructed by intercepting or filling, wherein the filled vector is the average value of the low-latitude real-valued vectors corresponding to all location IDs; in this embodiment, the first 30 location IDs are intercepted to Build a trajectory matrix of fixed dimensions for each user;

S3. Build a prediction model

Construct a four-layer convolutional neural network, as shown in Figure 2, the input layer is a fixed-dimensional trajectory matrix; the convolutional layer sets three convolution kernels m*embedding_size, where m is 2, 3, 4. , the embedding size is the dimension of the low-latitude real-valued vector output by Node2vec, the number of each convolution kernel is 64; the pooling layer is k-max pooling, k represents the first k maximum values after convolution, in this implementation In the example, k=3; the output of the fully connected layer is input to the softmax function to obtain the probability distribution of the trajectory owner.

S4. Train the prediction model

S4.1, build a training set

The fixed-dimensional trajectory matrix set X of some users in the dataset and the corresponding one-hot category vector set Y are used as the training set, where X=[Vec_x ₁ ,Vec_x ₂ ,...,Vec_x _n ], Vec_x _n represents the first Trajectory matrix of n fixed dimensions of users, Y=[Vec_t ₁ , Vec_t ₂ ,...,Vec_t _n ], Vec_t _n represents the one-hot category vector corresponding to the nth owner, if the nth position is 1, Then the rest of the positions are all 0;

S4.2. Initialize the prediction model

The value of the weight matrix W _p of each convolution kernel in the initialized convolution layer is normal distribution, and its mean is 0 and the variance is 0.1; at the same time, the bias vector B _p of each convolution kernel in the convolution layer is initialized to 0.1, The number of elements of each bias vector is the number of neurons in the corresponding layer; the weight matrix of the initialized fully connected layer is W, and its dimension is [batch_size*k*Number of convolution kernels, number of categories], and at the same time initialize the full connection layer. The value of the bias vector B of the connection layer is 0.1, and the number of elements is the number of categories; among them, batch_size is a constant, the value is 64, p=1, 2, 3, p represents the number of convolution kernels in the convolution layer;

S4.3 The training set is input into the initialized prediction model, and the Adam algorithm is used to optimize the loss function, and the optimized loss function is as follows:

Wherein, N=batch_size, y _j is all elements in Vec_t _j , a _j is the output value of softmax function;

Then use the error back propagation BP algorithm to pass the error to the previous layer, update the weight matrix W _p , the bias vector B _p of the convolution layer and the weight matrix W and bias vector B of the fully connected layer, and after several iterations , to obtain the converged neural network model, thereby obtaining the trained prediction model;

S5, trajectory owner prediction

According to the method described in steps S1 and S2, construct the trajectory matrix of the fixed dimension of the user's position, and then input the constructed trajectory matrix into the already trained prediction model to obtain all the owner categories corresponding to the trajectory. The probability distribution of , and the index of the maximum value in the probability distribution is the owner number corresponding to the track.

Although the illustrative specific embodiments of the present invention have been described above to facilitate the understanding of the present invention by those skilled in the art, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, As long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious, and all inventions and creations utilizing the inventive concept are included in the protection list.

Claims (2)

1. a trajectory owner prediction method based on convolutional neural network, is characterized in that, comprises the following steps: (1), data preprocessing (1.1) Count the longitude and latitude of the historical location of all user trajectories in chronological order to form a longitude and latitude set, wherein, if a certain longitude and latitude appears repeatedly, it is only reserved once in the longitude and latitude set; Number each longitude and latitude in the longitude and latitude set from 1, and give a unique location ID; (1.2), in chronological order, use the location ID given in step (1.1) to replace the longitude and latitude of the historical location of the user track, then each user track is represented by a string of location IDs; at the same time, each user is represented by a unique integer ID Identify, so that the user and the user track can form the owner track in the form of [user ID, track (location ID, ..., location ID)]; (1.3) Form a directed unweighted graph G=<V, E> according to the owner trajectory, where V is the set of all location IDs, if a user appears from ID _i to ID _j , then <ID _i , ID _j > represents a directed edge, and all such edges constitute the edge set E of the graph G; (2), track representation (2.1), take the constructed directed unweighted graph G as the input, and learn the low-dimensional real-valued vector of each position ID in G through the Node2Vec algorithm; (2.2), slice the trajectory of each user in step (1.2) according to fixed time intervals, thereby dividing each user trajectory into several position ID sequences, and then use the user ID to own the cut position ID sequence. identification; (2.3), replace the position ID with the low-dimensional real-valued vector corresponding to the position ID to the indefinite-length trajectory after cutting, thereby generating the trajectory matrix of each user; Then, a fixed-dimensional trajectory matrix of each user is constructed by intercepting or filling to form a data set, where the filled vector is the average value of the low-latitude real-valued vectors corresponding to all location IDs; (3), build a prediction model Construct a four-layer convolutional neural network, whose input layer is a fixed-dimensional trajectory matrix; the convolutional layer sets three convolution kernels m*embedding_size, where m is a constant, and the embedding size is the low-dimensional real-valued vector output by Node2vec. Dimension; the pooling layer is k-max pooling, k represents the first k maximum values after convolution; the output of the fully connected layer is input to the softmax function to obtain the probability distribution of the track owner; (4), training prediction model (4.1), build a training set The fixed-dimensional trajectory matrix set X of some users in the dataset and the corresponding one-hot category vector set Y are used as the training set, where X=[Vec_x ₁ ,Vec_x ₂ ,...,Vec_x _n ], Vec_x _n represents the first Trajectory matrix of n fixed dimensions of users, Y=[Vec_t ₁ , Vec_t ₂ ,...,Vec_t _n ], Vec_t _n represents the one-hot category vector corresponding to the nth owner, if the nth position is 1, Then the rest of the positions are all 0; (4.2), initialize the prediction model The value of the weight matrix W _p of each convolution kernel in the initialized convolution layer is a normal normal distribution with a mean value of 0 and a variance of 0.1; at the same time, the bias vector B _p of each convolution kernel in the convolution layer is initialized is 0.1, the number of elements of each bias vector is the number of neurons in the corresponding layer; the weight matrix of the initialized fully connected layer is W, and its dimension is [batch_size*k*number of convolution kernels, number of categories], At the same time, the bias vector B value of the fully connected layer is initialized to 0.1, and the number of elements is the number of categories; among them, batch_size is a constant, p=1, 2, 3, p represents the number of convolution kernels in the convolution layer; (4.3) Input the training set into the initialized prediction model, use the Adam algorithm to optimize the loss function, and then use the error backpropagation BP algorithm to pass the error to the previous layer, update the weight matrix W _p , bias of the convolution layer The vector B _p , the weight matrix W of the fully connected layer, and the bias vector B, after several iterations, the converged neural network model is obtained, thereby obtaining the trained prediction model; (5), trajectory owner prediction According to the method described in steps (1) and (2), the latitude and longitude of the location of the user to be detected is used to construct a trajectory matrix of the fixed dimension of the user, and then the constructed trajectory matrix is input into the already trained prediction model, and the corresponding trajectory of the trajectory is obtained. The probability distribution of all owner categories, and the index of the maximum value in the probability distribution is the owner number corresponding to the track. 2. a kind of trajectory owner prediction method based on convolutional neural network according to claim 1, is characterized in that, the loss function that described Adam algorithm optimizes is as follows: Among them, N=batch_size, y _j is all the elements in Vec_t _j , and a _j is the output value of the softmax function.