CN115169526B - Base station representation learning method, system and storage medium based on deep learning - Google Patents

Abstract

The present invention provides a base station representation learning method, system and storage medium based on deep learning. The method includes: acquiring base station information, the base station information including unique identification code information of the base station, static feature data of the base station, and dynamic trajectory data related to the base station ; Determine the adjacency matrix between the base stations based on the acquired longitude, latitude and base station-related dynamic trajectory data, and construct the first base station relationship diagram based on the adjacency matrix; construct the first feature vector based on the static feature data of the base station and the relevant dynamic trajectory data of the base station set; construct the self-encoder neural network model, input the first base station relationship diagram and the first feature vector set to the encoder to obtain a base station representation vector set, and input the base station representation vector set to the decoder to obtain the reconstructed first Two base station relationship graphs and a second set of feature vectors. This method can efficiently realize the representation learning of the base station, so that the information of the base station can be better applied in various data mining processes.

Description

A base station representation learning method, system and storage medium based on deep learning

technical field

The present invention relates to the technical field of big data mining, in particular to a base station representation learning method, system and storage medium based on deep learning.

Background technique

The base station is an important mobile communication facility. Mobile communication users obtain services through the base station, so that they can meet the needs of making and receiving calls, sending and receiving short messages, and mobile Internet access. Therefore, base stations play an important role in mobile communication big data. For example, in the mobile communication trajectory data, the base station information in a certain trajectory point indicates the location range of the user at this moment.

With the development of deep learning technology, the application of deep learning models to mine mobile communication trajectory data has become the mainstream in recent years. However, in the existing methods, the base station identification code is usually mapped to the latitude and longitude information of the base station, and then the latitude and longitude data are mapped to grid codes, and the grid code embedding technology is used to transfer the mobile communication trajectory data to the neural network model for training. First of all, this preprocessing method cannot realize end-to-end mobile communication data mining; secondly, due to different base station standards, environments and other factors, the area ranges represented by different base stations are also different, and the characteristics of the base station itself are not consistent with the grid; Furthermore, compared with the grid, the base station also has the attribution information of the operating company, and there may be base stations of multiple operators within the same geographical area. How to learn representations of base stations and transform their features into low-dimensional space vectors, so that deep learning models can make better use of base station information in various mining tasks (such as trajectory prediction, abnormal base station detection), is an urgent technology to be solved question.

Contents of the invention

In view of this, the present invention provides a base station representation learning method, system and storage medium based on deep learning, so as to solve one or more problems existing in the prior art.

According to one aspect of the present invention, the present invention discloses a base station representation learning method based on deep learning, the method comprising:

Obtain base station information, the base station information includes the unique identification code information of the base station, base station static feature data and base station related dynamic trajectory data, the base station static feature data includes location area code, cell code, longitude, latitude, address, operator , direction angle, and POI information within a specific range, the dynamic track data related to the base station includes the time when the base station is visited, the time the user stays in the base station, and the number of visiting users of the base station;

Determining an adjacency matrix between base stations based on the acquired longitude and latitude of the base station and the relevant dynamic trajectory data of the base station, and constructing a first base station relationship diagram based on the adjacency matrix;

Constructing a first set of feature vectors based on the static feature data of the base station and the dynamic trajectory data related to the base station;

Constructing an autoencoder neural network model, inputting the first base station relationship graph and the first feature vector set to the encoder to obtain a base station representation vector set, and inputting the base station representation vector set to a decoder to obtain a reconstructed first Two base station relationship graphs and a second set of feature vectors.

In some embodiments of the present invention, the time when the base station is visited includes a time period when the base station is visited the most times and a time period when the base station is visited the least times;

The time that the user stays in the base station includes the average value and median value of the time that the user stays in the base station; and/or

The number of visiting users of the base station includes the total number of visiting users of the base station on working days and non-working days, the average number of visiting users per hour on working days and non-working days, the maximum number of visiting users on working days and non-working days, the number of visiting users on working days and The median number of visiting users during non-working days and the minimum number of visiting users during working days and non-working days.

In some embodiments of the present invention, each element in the adjacency matrix is 0 or 1; when the distance between the first base station and the second base station is less than a preset distance, and the first base station and the second base station When the total number of times that the base station is used as an adjacent track point is not less than the preset number, the corresponding element in the adjacency matrix is 1.

In some embodiments of the present invention, the method also includes:

build loss function;

Updating the weight parameters of the autoencoder neural network model.

In some embodiments of the present invention, the loss function is: Among them, V _i is the first eigenvector of the i-th base station, /> is the second eigenvector of the i-th base station, N is the total number of base stations, h _i is the representation vector of the i-th base station, λ is a hypermodel parameter, M is the total number of neighboring base stations of the i-th base station, h _j is The representation vector of the neighbor base station j of the i-th base station.

In some embodiments of the present invention, the decoder includes a self-attention module, and Q, K, and V of the self-attention module are all σ(W ^k h ^k-1 ); wherein, σ is a nonlinear activation function, Relu function can be selected; W ^k is a trainable weight coefficient matrix with a size of N×L, where N is the number of base stations, L is the size of h ^k-1 , and the initial value of W ^k is randomly initialized between 0 and 1 The real number of; h ^k-1 is the output of the k-1th layer graph neural network.

In some embodiments of the present invention, both the encoder and the decoder include a three-layer graph attention neural network layer.

In some embodiments of the present invention, the method also includes:

Set the number of iterations and learning rate of the autoencoder neural network model;

The autoencoder neural network model is trained.

According to another aspect of the present invention, a base station representation learning system based on deep learning is also disclosed, the system includes a processor and a memory, computer instructions are stored in the memory, and the processor is used to execute the instructions stored in the memory. computer instructions, when the computer instructions are executed by the processor, the system implements the steps of the method described in any one of the above embodiments.

According to yet another aspect of the present invention, a computer-readable storage medium is also disclosed, on which a computer program is stored. When the program is executed by a processor, the steps of the method described in any one of the above embodiments are implemented.

The base station representation learning method and system based on deep learning disclosed in the embodiment of the present invention obtains the representation vector of the base station through self-encoder neural network model learning based on the first base station relationship graph and the first feature vector, that is, projects the base station features to low-dimensional Space vector, so that the representation learning of the base station can be realized efficiently, so that the base station information can be better applied in various data mining processes.

Additional advantages, objects, and features of the present invention will be set forth in part in the following description, and will be partly apparent to those of ordinary skill in the art after studying the following text, or can be learned from the practice of the present invention. The objectives and other advantages of the invention may be realized and attained by the structure particularly pointed out in the written description and claims hereof as well as the appended drawings.

It will be understood by those skilled in the art that the objects and advantages that can be achieved by the present invention are not limited to the above specific ones, and the above and other objects that can be achieved by the present invention will be more clearly understood from the following detailed description.

Description of drawings

The drawings described here are used to provide further understanding of the present invention, constitute a part of the application, and do not limit the present invention. The components in the figures are not drawn to scale, merely illustrating the principles of the invention. For ease of illustration and description of some parts of the present invention, corresponding parts in the figures may be exaggerated, ie, may be made larger relative to other components in an exemplary device actually manufactured in accordance with the present invention. In the attached picture:

FIG. 1 is a schematic flowchart of a base station representation learning method based on deep learning according to an embodiment of the present invention.

FIG. 2 is a schematic structural diagram of a base station representation learning model based on deep learning according to an embodiment of the present invention.

FIG. 3 is a schematic diagram of a data update process of an encoder of a base station representation learning model according to an embodiment of the present invention.

Detailed ways

In order to make the purpose, technical solutions and advantages of the embodiments of the present invention more clear, the embodiments of the present invention will be further described in detail below in conjunction with the accompanying drawings. Here, the exemplary embodiments and descriptions of the present invention are used to explain the present invention, but not to limit the present invention.

Here, it should be noted that, in order to avoid obscuring the present invention due to unnecessary details, only the structures and/or processing steps that are closely related to the solution according to the present invention are shown in the drawings, while those related to the present invention are omitted. Invent other details that don't really matter.

It should be emphasized that the term "comprises/comprises/has" when used herein refers to the presence of a feature, element, step or component, but does not exclude the presence or addition of one or more other features, elements, steps or components.

In order to perform better representation learning on the base station and convert the features of the base station into low-dimensional space vectors, so that the deep learning model can better use the information of the base station in various mining tasks, the present invention provides a base station based on deep learning Represents a learning method, system and storage medium.

Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. In the drawings, the same reference numerals represent the same or similar components, or the same or similar steps.

FIG. 1 is a schematic flow chart of a base station representation learning method based on deep learning according to an embodiment of the present invention. As shown in FIG. 1 , the base station representation learning method includes at least steps S10 to S40.

Step S10: Obtain base station information, the base station information includes the unique identification code information of the base station, base station static characteristic data and base station related dynamic trajectory data, the base station static characteristic data includes location area code, cell code, longitude, latitude, address, The operator, direction angle, and POI information within a specific range. The dynamic trajectory data related to the base station includes the time when the base station is visited, the time users stay in the base station, and the number of users visiting the base station.

The base station information can be pre-saved in the base station information table or the corresponding database. Specifically, the information of the base station that needs to be learned can be obtained from the base station information table or the corresponding database. The unique identification code information refers to the ID information of the base station . Exemplarily, the first base station to the Nth base station can be respectively recorded as BS ₁ , BS ₂ ... BS _n , and BS ₁ represents the unique identification code of the base station 1; correspondingly, the base station static feature data of the base station 1 to the base station n can be _They are _denoted as _{e 1} , _{e 2} .

In this step, the location area code (LAC, Location Area Code) is used to identify different location areas, one location area may contain one or more cells, and the location area code is included in the location area identification code (LAI). The cell code (CI, Cell Identifier) is used as the unique identification code of the cell, and the cell code (CI) is combined with the location area identifier (LAI) to identify the cell covered by each base station in the network. The address refers to the specific location of the base station, such as No. XX, XX Street, XX District, XX City. The operator of the base station is such as China Mobile, China Unicom or China Telecom. In the base station information table in this embodiment, each operator can be represented by a corresponding code. The direction angle of the base station is determined by the orientation of the antenna in the base station. The POI information in a specific range refers to the specific information of the POI in a specific area around the base station with the base station as the center. The POI in this embodiment can specifically be Stores, hospitals, schools, gas stations, etc. It should be understood that the specific range may be set according to actual needs. For example, in an embodiment, the specific range may be limited to a circular area around the base station with a radius of k kilometers.

The visited time of the base station may include a time period when the base station is visited the most times and a time period when the base station is visited the least times. At this time, the visited data of the base station can be obtained first, for example, the number of times each base station is visited within a day, the specific time of each visit, etc.; and then the statistics of each time interval (time period) of each base station within a day The times of visits; the time interval can be hours, for example, and at this time, the times of visits of each base station in each hour of a day are counted. After counting the total number of times that the base station is visited in each hour of the day, it is further possible to determine the time period when the base station is visited the most and the least in a day, and the time period when the base station is visited the most can also be understood as The base station is visited most frequently during this time period.

The time that the user stays in the base station may specifically include an average value and a median value of the time that the user stays in the base station. Before determining the average value and median value of the time that users stay in the base station, the total time that each user stays in the base station within a preset time period (such as one day) and the total number of times that the base station is visited within a preset time period can be obtained first. The number of users, and then based on the sum of the total time that each user stays in the base station and the total number of users, the average and median values of the time that users stay in the base station are calculated.

The number of visiting users of the base station includes the total number of visiting users of the base station on working days and non-working days, the average number of visiting users per hour on working days and non-working days, the maximum number of visiting users on working days and non-working days, the maximum number of visiting users on working days and non-working days The median number of accessing users per hour per day and the minimum number of accessing users per hour on working days and non-working days. In this embodiment, the number of visiting users of the base station includes those on working days and those on non-working days, because it is considered that the difference between the office space and the entertainment place covered by the base station has a certain impact on the characteristics of the base station, for example, the coverage of office space on weekdays The number of visitors to base stations with more places on weekdays will be greater than that on non-workdays. It is not difficult to understand that the status of base stations with more entertainment places is just the opposite. Therefore, obtaining the relevant information of visiting users on working days and non-working days at the same time can better extract the semantic features of the base station. Exemplarily, whether it is a working day or a non-working day, the number of users who visit the base station every hour in a day can be obtained first, and then the total number of users who visit the base station in a day can be counted, and finally the number of users who visit the base station in each hour of the day can be compared The number of users can determine the time period with the largest and least number of users accessing the base station in the day and the specific number of users in each time period, and based on the total number of users accessing the base station in the day obtained from statistics, the number of users in the day can be calculated. The average number of hourly visiting users.

Step S20: Determine an adjacency matrix between base stations based on the obtained longitude and latitude of the base stations and the dynamic trajectory data related to the base stations, and construct a first base station relationship graph based on the adjacency matrix.

In this step, in order to further determine the adjacency matrix, the adjacency matrix of the base station can be recorded as M, and each element in the matrix M can be 0 or 1. Specifically, when the distance between the first base station and the second base station is less than the preset distance, and the total number of times the first base station and the second base station are adjacent trajectory points is not less than the preset number, the adjacency matrix The corresponding element is 1, otherwise, the elements in the adjacency matrix M are recorded as 0. Exemplarily, when the distance between the third base station and the fourth base station is greater than the preset distance, and the total number of times the third base station and the fourth base station are adjacent track points is not less than the preset number; or when the third base station and the fourth base station When the distance between the fourth base stations is greater than the preset distance, and the number of times the third base station and the fourth base station are adjacent track points is less than the preset number of times; the corresponding elements in the adjacency matrix M are all recorded as 0. It should be understood that the preset distance and the preset times can be limited according to actual needs; in one embodiment, the preset distance can be limited to 3 kilometers, and the preset times can be set to 20 times.

Since each element in the adjacency matrix reflects the adjacent base stations of the base station and the distance between the base station and the adjacent base station, in this step, the base station is used as the vertex, and the distance between the base station and the adjacent base station is used as the side to construct the first base station relationship picture.

Step S30: Construct a first set of feature vectors based on the static feature data of the base station and the dynamic track data related to the base station.

In this step, the eigenvectors of each base station may be determined first, and then the first combination of eigenvectors is established based on the determined eigenvectors of each base station. The feature vector composition of the base station can include: area code, cell code, longitude, latitude, address, operator, direction angle, POI information within a specific range, time period when the base station is visited the most times, and time when the base station is visited the least times segment, the average and median value of the time that users stay in the base station, the total number of visiting users of the base station on working days and non-working days, the average number of visiting users on working days and non-working days, and the hourly visits on working days and non-working days The maximum number of users, the median number of visiting users per hour on working days and non-working days, and the minimum number of visiting users per hour on working days and non-working days. Wherein, the POI information within a specific range may specifically be the total number of POIs of various types within the preset range around the base station, such as the number of shopping places, the number of entertainment places, the number of dining places, etc.; and in this embodiment, In addition to the content listed above, the base station feature vector may also include more or less content.

Step S40: Construct a self-encoder neural network model, input the first base station relationship graph and the first feature vector set to the encoder to obtain a base station representation vector set, and input the base station representation vector set to a decoder to obtain a re- Structured second base station relationship graph and second feature vector set.

This step is to construct the self-encoder neural network model to learn the representation of the base station. The self-encoder neural network model includes an encoder module and a decoder module. The input of the encoder module is the first base station relationship graph and the first feature vector set , and the output of the encoder is the set of vectors obtained from representation learning (base station representation vector set); the input of the decoder is the set of base station representation vectors output by the encoder, and the output of the decoder is the reconstructed second base station relationship diagram and the first A collection of two eigenvectors.

In addition, in order to obtain a more ideal autoencoder neural network model, iterative training is further performed on the autoencoder neural network model. Specifically, the base station representation learning method further includes the following steps: constructing a loss function; and updating weight parameters of the autoencoder neural network model. The loss function is used to measure the degree of inconsistency between the real value and the predicted value learned by the self-encoder neural network model. The smaller the loss function, the higher the robustness of the model. Specifically, when building an autoencoder neural network model, it is also necessary to set relevant hyperparameters, such as setting the number of iterations and learning rate of the autoencoder neural network model; after setting the number of iterations and learning rate of the autoencoder neural network model Afterwards, the autoencoder neural network model can also be trained based on the set number of iterations and learning rate. The data update process of the encoder of the base station representation learning model is shown in Figure 3.

In the present invention, the training and optimization objectives of the autoencoder neural network model are: the cost of reconstructing the base station feature vector is minimized, and the similarity of the standard vectors between the base station and its neighboring base stations is maximized. As an example, the constructed loss function is Among them, V _i is the first eigenvector of the i-th base station, /> is the second eigenvector of the i-th base station, N is the total number of base stations, h _i is the representation vector of the i-th base station, λ is a hypermodel parameter, M is the total number of neighboring base stations of the i-th base station, h _j is The representation vector of the neighbor base station j of the i-th base station.

In an embodiment of the present invention, the decoder includes a self-attention module, and Q (query), K (key), and V (value) of the self-attention module are all σ(W ^k h ^k-1 ); Among them, σ is a nonlinear activation function, and the Relu function can be selected; W ^k is a trainable weight coefficient matrix with a size of N×L, where N is the number of base stations, L is the size of h ^k-1 , and the initial value of W ^k is random Initialized as a real number between 0 and 1; h ^k-1 is the output of the k-1th layer graph neural network. In this embodiment, both the encoder and the decoder can include a three-layer graph attention neural network layer, and k can take values of 1, 2, and 3 at this time. It can be understood that including the three-layer graph attention neural network layer in both the encoder and the decoder is only a preferred example, which can be changed according to specific application scenarios.

Fig. 2 is a schematic structural diagram of a base station representation learning model based on deep learning according to an embodiment of the present invention. As shown in Fig. 2 , the original relationship diagram of the base station is constructed first, and the original relationship diagram of the base station and the corresponding base station feature vector set are input to the encoder A base station characterization vector is obtained, and then the base station characterization vector is input to a decoder to obtain a base station reconstruction relationship graph and a corresponding base station reconstruction feature vector set.

Correspondingly, the present invention also discloses a base station representation learning system based on deep learning, the system includes a processor and a memory, the memory stores computer instructions, and the processor is used to execute the computer instructions stored in the memory , when the computer instructions are executed by the processor, the system implements the steps of the method described in any one of the above embodiments.

Exemplarily, the base station representation learning system specifically performs the following steps: base station data acquisition, base station relationship graph construction, graph neural network construction, and base station representation learning. Specifically, base station data acquisition includes base station code identification basic data acquisition, base station static feature data acquisition, and base station dynamic feature data acquisition; base station static feature data includes base station geographic coordinates, longitude and latitude data, base station operator, base station direction angle, base station coverage, etc. Information; base station static data can be obtained from text files and databases. The dynamic feature data of the base station includes the dynamic trajectory data related to the base station, and the dynamic trajectory data related to the base station can be obtained from databases, message queues, and text files.

The construction of the base station relationship graph is to construct the base station relationship graph based on the base station data. Each vertex in the base station relationship graph represents a base station, and the base stations with correlation are connected with an edge. The base station correlation may depend on the distance between the base stations and the dynamic switching relationship between the base stations.

The construction of the graph neural network is to build a self-supervised graph neural network model, including an encoder and a decoder. The encoder is used to learn the base station representation vector from the base station relationship graph, and the decoder is used to compare the base station graph data from the learned base station representation vector. Perform reconstruction and perform self-supervised learning by optimizing the reconstruction loss.

The base station representation learning is based on the base station relationship graph and base station data, setting model hyperparameters, training the constructed graph neural network model, and learning the base station representation vector.

The above method will be described below in conjunction with a specific example. However, it should be noted that this specific example is only for better illustrating the present application, and does not constitute an improper limitation to the present application.

In this embodiment, a set of base stations D is obtained first, D={BS ₁ , BS ₂ ,...,BS _n }, where BS _i represents the unique identification code of base station i; n represents the total number of base stations. Then obtain the static feature data E of the base station, E={(BS ₁ ,e ₁ ),(BS ₂ ,e ₂ )...(BS _n ,e _n )}, where e _i represents the static feature vector of base station i corresponding to BS _i , usually including location area code (LAC, Location Area Code), cell code (CI, Cell Identifier), longitude, latitude, address, operator, azimuth, coverage and other information. In addition, according to the latitude and longitude information of the base station, the Baidu map service interface can be used to obtain POI (Point Of Interest) information in a circular area with the base station as the center and a radius of k kilometers. The specific value of K can be 3, then what is obtained at this time is the POI information in a circular area with the base station as the center and a radius of 3 kilometers.

After the static feature data of the base station is obtained, the dynamic trajectory data related to the base station is further obtained. Further base station dynamic characteristic data can be obtained based on the relevant dynamic trajectory data of the base station, and the dynamic characteristic data of the base station is recorded as F, then F={(BS ₁ ,f ₁ ),(BS ₂ ,f ₂ ),...,(BS _n ,f _n )}, where f _i represents the dynamic feature vector of base station i corresponding to BS _i . The base station dynamic feature vector usually includes information such as access time, dwell time, and number of access users. In this embodiment, the visit time includes: the hour when the base station is most frequently visited, and the hour when the base station is least frequently visited; the dwell time includes: the average value and median value of the user's stay time in the base station; the number of visiting users Including: the total number of visiting users of the base station on working days and non-working days, the average number of visiting users per hour on working days and non-working days, the maximum number of visiting users on working days and non-working days, and the hourly visiting users on working days and non-working days The median value of the number and the minimum number of access users per hour on weekdays and non-workdays, etc.

Further, the adjacency matrix M between the base stations is calculated based on the latitude and longitude data of the base stations obtained in the above steps and the dynamic characteristic data of the base stations, and M _{i, j} = 1 represents the base station i corresponding to BS _i to the base station _j corresponding to BS j The distance is less than k kilometers, and the number of times that base station BS _i and base station BS _j appear in the adjacent track points of the same track in the track data is not less than n times; otherwise M _{i, j} =0; in this implementation example, k, The values of n are 3 and 20 respectively. Further, based on the adjacency matrix M obtained in this step, a first base station relationship graph G is constructed.

At the same time, based on the obtained base station static characteristic data and the relevant dynamic trajectory data of the base station, the first characteristic vector set V of the base station is constructed. The first characteristic vector set V of a base station includes: the code of the operator to which the base station belongs; the base station location area code; is the center of the circle, the number of various types of POIs in a circular area with a radius of k kilometers; the hour when the base station is most frequently visited; the hour when the base station is least frequently visited; the average and median stay time of users in the base station Numerical value; the total number of visiting users of the base station on working days and non-working days, the average number of visiting users per hour on working days and non-working days, the maximum number of visiting users on working days and non-working days, the number of visiting users per hour on working days and non-working days value and the minimum number of access users per hour on weekdays and non-workdays.

Based on the determined first base station relationship graph G and the first feature vector V (V∈R ^N*d , N represents the number of base stations, and d represents the dimension of the first feature vector of the base station), construct a graph attention autoencoder neural network model . The model includes an encoder and a decoder. The input of the encoder is the first base station relationship graph G and the first feature vector set V of the vertices (ie base stations) in the graph G, and the output of the encoder is the vector obtained by base station representation learning. Set h (h∈R ^N*d′ , N is the number of base stations, d′ is the dimension of the learned base station representation vector, and the value is 128 in this embodiment). The input of the decoder is the vector set h obtained by base station representation learning, and the output is the reconstructed second base station relationship graph and the second set of eigenvectors /> There are two training objectives of the model, one is to minimize the loss cost of base station feature vector reconstruction, /> The other is to maximize the standard vector similarity between the base station and its neighbor base stations, /> The overall loss function is expressed as: where N is the total number of base stations, h _i is the representation vector of the i-th base station, M is the total number of neighbor base stations of the i-th base station, _hj is the representation vector of the i-th base station’s neighbor base station j, and λ is the model hyperparameter , λ is 0.5 in this implementation example.

In this embodiment, the number of layers of the encoder and the decoder is set to 3, and the information of the graph neural network of the k-1th layer is passed to the kth layer through the self-attention mechanism. Specifically, the query Q, key K, and value V in attention are set as: Q ^k , K ^k , V ^k = σ(W ^k h ^k-1 ), and the weights of base station i and other base stations are: The representation vector of base station i is updated as: Among them, Nei(i) represents the set of neighbor base stations of base station i, and j represents the neighbor base stations of base station i.

Further, set relevant hyperparameters to train the above-mentioned graph self-attention encoder neural network model, where the base station representation vector dimension is set to 128, that is, the hidden layer size is 128, the number of codec layers is set to 3, and the model The number of training iterations is set to 100, and the learning rate is set to 10 ^-3 .

Through the above-mentioned embodiments, it can be found that the deep learning-based base station representation learning method and system of the present invention, based on the first base station relationship graph and the first feature vector, obtains the base station representation vector through self-encoder neural network model learning, that is, the base station feature projection To a low-dimensional space vector, so that the representation learning of the base station can be realized efficiently, so that the base station information can be better applied in various data mining processes.

In addition, the invention also discloses a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the steps of the method described in any one of the above embodiments are realized.

Those of ordinary skill in the art should understand that each exemplary component, system and method described in conjunction with the embodiments disclosed herein can be implemented by hardware, software or a combination of the two. Whether it is implemented in hardware or software depends on the specific application and design constraints of the technical solution. Skilled artisans may use different methods to implement the described functions for each specific application, but such implementation should not be regarded as exceeding the scope of the present invention. When implemented in hardware, it may be, for example, an electronic circuit, an application specific integrated circuit (ASIC), suitable firmware, a plug-in, a function card, or the like. When implemented in software, the elements of the invention are the programs or code segments employed to perform the required tasks. Programs or code segments can be stored in machine-readable media, or transmitted over transmission media or communication links by data signals carried in carrier waves. "Machine-readable medium" may include any medium that can store or transmit information. Examples of machine-readable media include electronic circuits, semiconductor memory devices, ROM, flash memory, erasable ROM (EROM), floppy disks, CD-ROMs, optical disks, hard disks, fiber optic media, radio frequency (RF) links, and the like. Code segments may be downloaded via a computer network such as the Internet, an Intranet, or the like.

It should also be noted that the exemplary embodiments mentioned in the present invention describe some methods or systems based on a series of steps or devices. However, the present invention is not limited to the order of the above steps, that is, the steps may be performed in the order mentioned in the embodiment, or may be different from the order in the embodiment, or several steps may be performed simultaneously.

In the present invention, features described and/or exemplified for one embodiment can be used in the same or similar manner in one or more other embodiments, and/or can be combined with features of other embodiments or replace other Features of the implementation.

The above descriptions are only preferred embodiments of the present invention, and are not intended to limit the present invention. For those skilled in the art, various modifications and changes may be made to the embodiments of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims (8)

1. A method for deep learning-based base station representation learning, the method comprising:

acquiring base station information, wherein the base station information comprises unique identification code information of a base station, base station static characteristic data and base station related dynamic track data, the base station static characteristic data comprises position area codes, cell codes, longitudes, latitudes, addresses, affiliated operators, direction angles and POI information in a specific range, and the base station related dynamic track data comprises the accessed time of the base station, the stay time of users in the base station and the number of access users of the base station;

determining an adjacency matrix between base stations based on the acquired longitude and latitude of the base stations and the related dynamic track data of the base stations, and constructing a first base station relation diagram based on the adjacency matrix; the adjacency matrix is used for representing the distance between the base stations and the dynamic switching relation;

constructing a first feature vector set based on the base station static feature data and the base station related dynamic track data;

constructing a self-encoder neural network model, inputting the first base station relation diagram and the first characteristic vector set into an encoder to obtain a base station expression vector set, and inputting the base station expression vector set into a decoder to obtain a reconstructed second base station relation diagram and second characteristic vector set;

each element in the adjacency matrix is 0 or 1;

when the distance between the first base station and the second base station is smaller than a preset distance, and the total number of times of the first base station and the second base station serving as adjacent track points is not smaller than the preset number of times, the corresponding element in the adjacent matrix is 1;

the decoder includes self-attention modules Q, K, V each of which is σ (W ^k h ^k-1 ) The method comprises the steps of carrying out a first treatment on the surface of the Wherein σ is a nonlinear activation function; w (W) ^k Is a trainable weight coefficient matrix with the size of N multiplied by L, wherein N is the number of base stations, and L is h ^k-1 Size W of (2) ^k The initial value is randomly initialized to be a real number between 0 and 1; h is a ^k-1 Is the output of the neural network of the k-1 layer.

2. The deep learning-based base station representation learning method according to claim 1, wherein the accessed time of the base station includes a time period in which the number of times the base station is accessed is the largest and a time period in which the number of times the base station is accessed is the smallest;

the time for the user to stay in the base station comprises the average value and the median value of the time for the user to stay in the base station; and/or

The access user quantity of the base station comprises total number of access users of the base station on working days and non-working days, average number of access users of the base station on working days and non-working days, maximum number of access users of the base station on working days and non-working days, median number of access users of the base station on working days and non-working days and minimum number of access users of the base station on hours.

3. The deep learning-based base station representation learning method of claim 1, further comprising:

constructing a loss function;

and updating weight parameters of the self-encoder neural network model.

4. A deep learning based base station representation learning method according to claim 3, characterized in that the loss function is:wherein V is _i For the first eigenvector of the i-th base station,the second eigenvector of the ith base station, N is the total number of base stations, h _i Is the expression vector of the ith base station, lambda is the supermodel parameter, M is the total number of neighbor base stations of the ith base station, h _j Is a representation vector of a neighbor base station j of the i-th base station.

5. The deep learning based base station representation learning method of any one of claims 1 to 4, wherein the encoder and the decoder each comprise a three-layer graph attention neural network layer.

6. The deep learning based base station representation learning method of claim 4, further comprising:

setting the iteration times and the learning rate of the self-encoder neural network model;

training the self-encoder neural network model.

7. A deep learning based base station representation learning system comprising a processor and a memory, wherein said memory has stored therein computer instructions for executing the computer instructions stored in said memory, which system when executed by the processor implements the steps of the method according to any of claims 1 to 6.

8. A computer readable storage medium, on which a computer program is stored, characterized in that the program, when being executed by a processor, carries out the steps of the method according to any one of claims 1 to 6.