CN117635218A - Business district flow prediction method based on six-degree separation theory and graph annotation network - Google Patents

Abstract

The invention provides a business turn flow prediction method based on a six-degree separation theory and a graph annotation network, which comprises the following steps: s1, acquiring business turn flow data; s2, constructing a self-adaptive business turn flow diagram structure based on a six-degree separation theory; s3, extracting time features based on a linear gating convolution attention unit; s4, extracting spatial features based on the attention of the multi-head graph; s5, outputting a prediction result through the full connection layer, and completing establishment of a business turn flow prediction model; s6, training a business turn flow prediction model; s7, calling a trained business district flow prediction model to conduct flow prediction. The invention solves the problems of special business circle characteristics of business circle flow prediction and node long-distance space correlation, time lag and dynamic coupling caused by business marketing activities, and has obvious difference with space-time characteristic extraction in other flow prediction processes.

Description

Business district traffic prediction method based on six degrees of separation theory and graph attention network

Technical field

The invention belongs to the field of business prediction technology, and in particular relates to a business district traffic prediction method based on six degrees of separation theory and graph attention network.

Background technique

Business district traffic forecast plays an important role in shopping mall operations and management. Through accurate traffic prediction, shopping malls can customize different business strategies. Traffic prediction is an indispensable part of the smart shopping mall management system. The existing traffic prediction models are roughly divided into, and prediction models based on deep neural networks. Most prediction models based on linear statistical theory and machine learning implement static modeling predictions, considering information on a separate spatial or time scale, but cannot mine the dynamic dependencies between multi-variable time series data. Based on deep neural networks The prediction model has made significant progress in traffic prediction. Among them, graph neural network is more suitable for modeling complex relationships in node networks. They effectively capture the complex relationships between nodes with the help of graph structures, enable nodes to aggregate and disseminate information by iteratively updating node features, and show outstanding achievements in traffic prediction.

There are three types of graph model building methods: domain knowledge-based methods, text data-based methods, and data-based methods.

(1) The graph model construction method based on domain knowledge represents the nodes in the data as nodes of the graph, and determines the connection relationship between the nodes based on the knowledge of domain experts or prior information; or constructs domain knowledge as entity relationships. Tuples, construct entity-relationship graphs, in which entities serve as nodes, and relationships between entities are represented by edges. The weights of edges can reflect the strength or relevance of the relationships; both construction methods rely on domain expert knowledge.

(2) The graph model construction method based on text data converts the information in the text data into a graph structure, such as creating a word co-occurrence graph, in which words in the text are regarded as nodes, and the co-occurrence relationships between words are represented by edges. means that the edge weight can represent the co-occurrence frequency.

(3) Data-based graph model construction methods include k-nearest neighbor graph and radius neighbor graph. k-nearest neighbor graph construction represents each data point as a node of the graph and calculates the distance or similarity between nodes. , select the nearest k nodes to establish edge connections; the radius neighbor graph also represents data points as nodes, but determines whether the nodes are connected based on the predefined distance radius.

These methods provide the underlying graph structure for a variety of tasks by capturing relationships between data points. The data-based construction method has the advantages of automation, data-driven, interpretability, and wide applicability, making this method a versatile and powerful tool for building graph models. Because they extract information directly from the data and do not rely on domain knowledge or interpretation of text data, they are universally applicable to various data types and fields, and are suitable for large-scale data processing and machine learning tasks.

However, existing data-based graph model construction methods and graph neural network-based traffic prediction solutions still have limitations and are difficult to apply to business district traffic prediction in real life. The main reasons are:

1. When constructing the graph model, the existing method only considers the real location distribution between nodes, but does not take into account the impact of daily visible marketing models and shopping mall activities in business district traffic prediction - between each shopping mall and store location. The spatial dependence of is highly dynamic rather than static.

2. Existing solutions have an over-smoothing problem, making it difficult to capture long-distance spatial correlations - there are multiple shopping malls in a business district, there is a certain similarity between shopping malls, and two distant store locations may have similar traffic. Structural, spatial dependence is long-distance.

3. Existing methods do not consider the impact of time delays that may occur in the propagation of spatial information between locations. For example, when a location starts a product promotion, it takes a while (delay) to affect traffic conditions between stores.

Contents of the invention

The purpose of the embodiments of the present invention is to provide a business district traffic prediction method based on six degrees of separation theory and graph attention network, dynamically adjust the mapping method according to the business district, capture long-distance spatial correlation, and introduce linear gated convolution. Attention unit realizes business district traffic prediction.

In order to solve the above technical problems, the technical solution adopted by the present invention is a business district traffic prediction method based on six degrees of separation theory and graph attention network, which includes the following steps:

S1. Collect business district traffic data;

S2. Construct an adaptive business district traffic graph structure based on six degrees of separation theory;

S3. Extract temporal features based on linear gated convolution attention unit;

S4. Extract spatial features based on multi-head graph attention;

S5. Output the prediction results through the fully connected layer to complete the construction of the business district traffic prediction model;

S6. Business district traffic prediction model training;

S7. Call the trained business district traffic prediction model to predict traffic.

Further, the business district traffic data collected by S1 includes the business district range, store location, collection frequency, and collection time period.

Furthermore, the construction process of the adaptive business district traffic diagram structure based on the six degrees of separation theory in S2 is as follows:

S21. Calculate the correlation coefficient of any two store node variables to obtain the affinity matrix S; normalize the affinity matrix S, then any element in the matrix The value is between 0 and 1; arrange the upper triangular matrix elements in the affinity matrix S in descending order; when/> When , change the corresponding position of the adjacency matrix A/> Set to 1, and the rest to 0, to obtain the adjacency matrix A;/> is the threshold;

S22. Calculate the corresponding network average consistency for the adjacency matrix A. and the network average clustering coefficient CC , finally forming; the intersection point of the two curves/> A balance between efficiency and network redundancy;/> for/> The value of the network radius at the intersection point with the CC curve;

S23. Set the balance point value as the optimal threshold, taking the distance to be greater than or equal to/> Node pairs are connected to form a business district traffic node graph structure.

Further, the S3 is based on the linear gated convolution attention unit to extract temporal features as follows: first, replace the self-attention mechanism in the Transformer model with local context-sensitive convolution attention to form a gated attention unit, which is intended to be Combined with the time dimension characteristics of time series data; the input is the time length of Historical time series data, the input time series data is divided into blocks, and accurate convolution attention is used within the block and fast linear attention is used across blocks to obtain a gated attention unit with linear complexity; then by The degree-gated attention unit performs temporal convolution on the input data and outputs the feature vector of the correlation between each timestamp.

Further, the S4 extracts spatial features based on multi-head graph attention specifically as follows:

Input the temporal feature vector obtained by S3 into the spatial convolution layer to calculate the feature vector with spatial characteristics. The specific calculation process is as follows:

Among them,/> Is a matrix splicing operation; for the same node, multi-head self-attention is calculated separately/> attention, and merged by splicing or averaging/> attention;/> For the first/> Layer node/> eigenvector,/> For the first/> Layer node/> The feature vector of , after going through the aggregation operation (that is, the operation in the formula) with the attention mechanism as the core, the output is a node/> new eigenvector/> , can be understood as different nodes/> Characteristic Expression, Chapter/> nodes/> , the first one after going through the attention mechanism/> node feature representation.

Refers to the head of attention,/> Represents the attention coefficient,/> Represents node/> degree,/> ,/> is a learnable parameter,/> Is the first/> Vector representation of nodes, /> Indicates that the index is/> Middle/> Node pair/> The attention weight of the node;/> is the activation function,/> Is the first/> Vector representation of nodes;

Attention coefficient The calculation method is as follows:

Among them/> is a learnable parameter, No./> Attention coefficient of size,/> is an exponential function,/> Is a node/> eigenvector.

Further, the S5 is specifically: fusing temporal feature extraction and spatial feature vector through a fully connected layer and outputting a length of Prediction results/> ;wherein,/> Represents the node index, /> Indicates time,/> Shown in/> Time node/> predicted value.

Further, the process of training the S6 model is specifically:

S61 trains the model through the mean square error loss function. The loss function is:

Among them/> is the total number of samples in the time series used to calculate the loss,/> is the true value,/> is the predicted value;

S62. Use the Adam optimization algorithm to calculate the gradient of the network error for each weight parameter in backpropagation, and obtain new weights through the parameter update process; iteratively calculate the model weight until a predetermined small loss is reached, and the best prediction value is obtained.

Further, the S7 calls the trained business district traffic prediction model to perform traffic prediction as follows:

S71. Start real-time data collection, maintain a sampling frequency of once every 5 minutes, and input the collected data into the graph structure obtained by S2;

S72. Call the model built in S6 to output the predicted traffic value.

The beneficial effects of the present invention are

Aiming at the problem that business districts with complex commercial attributes are susceptible to traffic being affected by marketing and activities, the present invention proposes an adaptive business district traffic prediction method based on the idea of six degrees of separation, which solves the problems caused by marketing activities in commercial scenarios. Dynamic coupling relationships, over-smoothing problems caused by the similarity of shopping malls in business districts, and traffic lag problems in marketing activities. The present invention can be directly applied to different business districts, and there is no need to manually adjust the graph structure of the spatial relationship of traffic nodes in the business districts due to differences in business districts.

Description of drawings

In order to explain the embodiments of the present invention or the technical solutions in the prior art more clearly, the following will briefly introduce the drawings needed to describe the embodiments or the prior art. Obviously, the drawings in the following description are only These are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without exerting creative efforts.

Figure 1 is a flow chart of the detection method of the present invention;

Figure 2 is a flow chart of the adaptive adjacency matrix generation algorithm based on the idea of six degrees of separation;

Figure 3 is a structural diagram of the business district traffic prediction model based on the idea of six degrees of separation.

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.

As shown in Figure 3, the present invention provides a business district traffic prediction method based on six degrees of separation theory and graph attention network, which specifically includes the following seven steps. The first step is data statistics, the second step is the construction of a graph model of business district traffic, and the third step The fifth step is the construction of the prediction model, the sixth step is the model training strategy, and the seventh step is the deployment and application of this method. Details are as follows:

Step One: Traffic Data Statistics

First, formulate a detailed business district traffic data collection plan, including selecting the business district scope, store location, collection frequency, and collection time period. Deploy business district traffic data collection equipment at the entrance of the store to ensure that the location and angle of the equipment are suitable for capturing traffic conditions. Start the equipment as planned, start recording traffic data, and establish a data storage and management system. During the entire process, data is regularly cleaned and verified, timestamps are added, and data is backed up to ensure data security and privacy.

Step 2: Construct an adaptive business district traffic diagram model based on six degrees of separation theory

Based on the six degrees of separation theory and the characteristics of high marketing message dissemination efficiency and local aggregation of shops in real business districts, combined Radius network construction technology is used to construct business district traffic graph models. Figure 2 is a schematic flow chart of this algorithm, where S is the affinity matrix composed of the correlation between any two traffic nodes, and A represents the affinity matrix based on /> The node adjacency matrix of the graph structure constructed by the radius network construction technology, CC (A) quantifies the degree of connection between nodes in the complex network, and h (A) is used to measure the global transmission capacity of the complex network. In the statistical chart, the ordinate represents the values of CC(A) and h(A), and the abscissa represents different radius values.

Based on the six degrees of separation theory, connections can be established between nodes in a business district through no more than six intermediate levels of contact, and marketing messages can be spread through a very small number of nodes. At the same time, the distribution of shops in the business district is divided into areas according to categories, and clustering can be achieved according to customized distances (such as correlation between shops, actual distance between shops, etc.), so according to Radius network construction technology can realize graph construction between business district nodes. During the composition process, the threshold/> The choice is very important, choose the larger When , the present invention will obtain an extremely sparse random network. When/> As it gradually decreases, the network connections will gradually become denser, and its characteristics will begin to evolve toward a fully connected regular network. Therefore, this invention uses the sparse network average consistency ( h) and the clustering coefficient ( CC) to quantitatively calculate the propagation efficiency and sparsity of the graph structure to find the balance between the two, thereby determining the threshold/> . The specific construction process of the graph structure is as follows:

First, calculate the correlation coefficient of any two store node variables to obtain the affinity matrix S. After normalization, each element in the matrix The value is between 0 and 1.

Then the upper triangular matrix elements in the affinity matrix S are arranged in descending order; threshold Starting from 1, take the interval/> ,Right now .

like , then the corresponding position of the adjacency matrix A/> Set to 1 and the rest to 0, we get the adjacency matrix A.

Will gradually decreases to 0, and a total of/> A graph structure.

For the obtained adjacency matrix, calculate the corresponding network average consistency respectively. and the network average clustering coefficient CC , finally forming/> and CC curve. /> is an indicator to measure the efficiency of network communication, and CC is an indicator to measure the degree of connection between network nodes. Our goal is to construct a graph structure that is both efficient and informative, and the intersection of the two curves/> is the balance point between efficiency and network redundancy, and is therefore the optimal threshold for constructing the adjacency matrix.

Then, we define the equilibrium point as value as a threshold, the distance is greater than or equal to/> Node pairs are connected to form a business district traffic node graph structure. The business district traffic diagram model construction method of the present invention will adaptively and dynamically adjust the construction parameters according to the distribution of shopping malls and shops in different business districts. This method reduces the workload when the method is applied in reality, and the implementer does not need to set parameters based on experience. Then conduct observation experiments to select the optimal parameters, which is time-consuming and labor-intensive and has no theoretical support.

Step 3: Temporal feature extraction based on linear gated convolutional attention unit

This step is the temporal feature extraction layer of the model. This layer is the beginning of the model. The input is the time length of historical time series data, and at the same time input the adjacency matrix constructed in the second step.

Since RNN, LSTM and other networks currently used to extract nonlinear temporal characteristics are difficult to capture long-term dependencies, the Transformer model performs point-by-point dot product self-attention on time series, ignoring contextual information and being insensitive to the impact of surrounding commercial marketing activities. . Therefore, the present invention replaces the self-attention mechanism in Transformer with local context-sensitive convolution attention to form a gated attention unit to fit the time dimension characteristics of time series data. Due to the quadratic complexity of the gated attention unit, commercial district traffic prediction requires the use of long commercial district traffic data, resulting in a huge amount of calculation. To address this problem, the present invention divides the input time series data into blocks, uses precise convolutional attention within the block and uses fast linear attention across blocks, thereby obtaining a gated attention unit with linear complexity.

In detail, this layer uses a linear gated convolution attention unit to perform temporal convolution and adaptively capture the correlation of data between timestamps. CA focuses attention on local context changes, thereby learning more information. Many features. After passing through this temporal convolutional layer, the output is a feature vector that captures the correlation between each timestamp. The structure of the linear gated convolutional attention unit is shown in Figure 3, where CA represents the convolutional attention module, ACA represents the convolutional attention calculated based on the input, and dense represents the dense connection layer to weight and integrate information. Xt*UT+bT, Xt*WT+cT are fast linear attention modules, σ represents the activation function, Represents cross product.

The time feature extraction method proposed by the present invention can solve the unique business district characteristics of business district traffic prediction and the long-distance spatial correlation, time lag, and dynamic coupling problems of nodes caused by commercial marketing activities, and is different from the time and space problems in the traffic prediction process of other existing technologies. There are obvious differences in feature extraction.

Step 4: Spatial feature extraction based on multi-head graph attention

In order to fully utilize the information of nodes and their neighbors, this method uses multi-head graph attention to build spatial convolutional layers. Multi-head graph attention can model long-term dependencies and short-term dependencies at the same time, calculate attention in different subspaces in parallel, and complete multi-step long-distance prediction. The extracted temporal feature vector obtained through the third step is input into the spatial convolution layer, and the feature vector with spatial characteristics is obtained through calculation. The specific calculation process is as follows:

Among them,/> is a matrix splicing operation. For the same node, multi-head self-attention is calculated separately/> attention, and merged by splicing or averaging/> attention. /> For the first/> node/> characteristic variables,/> Refers to the head of attention,/> is the attention vector, /> Represents node/> degree,/> ,/> is a learnable parameter,/> Is the first/> Vector representation of nodes,/> Indicates that the index is/> Middle/> Node pair/> The attention weight of the node.

Multi-head self-attention can dig deeper into the potential of node data, allowing the model to better understand the characteristic meaning of nodes. The calculation of the attention coefficient is completed by a 2-layer multi-layer perceptron (MLP). The specific formula is as follows:

Among them/> is a learnable parameter/> is the activation function. /> No./> Attention coefficient of size,/> Represents the attention vector, /> is an exponential function,/> Is the node index, /> Is a node/> eigenvector.

Step 5: Fully connected layer.

After temporal feature extraction and spatial feature extraction, through a fully connected layer, the length is The final output/> . Among them,/> Represents the node index, /> Indicates time,/> expressed in Time node/> The predicted value of , here refers to all nodes.

Step 6: Business district traffic prediction model training.

After the model is built, model training begins. The model training strategy is as follows:

The mean square error loss function is used for model training optimization. The loss function is defined as:

Among them/> is the total number of samples in the time series used to calculate the loss,/> is the true value,/> is the predicted value.

According to the above framework, the Adam optimization algorithm is used to calculate the gradient of the network error for each weight parameter in backpropagation, and the new weight is obtained through the parameter update process. Model weights are calculated iteratively until a predetermined small loss is reached and the best predicted value is obtained. The reason for choosing Adam as the optimization algorithm is that it can design independent adaptive learning rates for different parameters. Most importantly, using Adam makes calculations more efficient. The overall training algorithm is as follows:

(1) Input multi-variable time series data X and adjacency matrix A;

(2) Initialize the model: determine the learning rate , batch size/> , number of iterations iter , initialization weight matrix W and bias matrix/> .

(3) Cycle training network:

A. Sample b data from the multi-variable time series data X each time until all the data is collected:

1. Calculation

2. Calculate the loss

3. Calculate the gradient and update the parameters using an optimization algorithm, where the learning rate is .

B.iter+1

(4) Complete the cycle and the model training is completed.

Step 7: Model deployment and traffic prediction

As shown in Figure 1, start real-time data collection, maintain a sampling frequency of once every 5 minutes, and input the collected data into the designed graph structure. Subsequently, the built and trained model is called, and the model outputs the predicted traffic value. This process will help managers better understand the dynamic changes in traffic and provide real-time decision support for business district management and operations.

Based on the six-degree separation theory, the present invention proposes an adaptive business district traffic graph model construction method, which completes the spatial modeling of traffic between different shops in the business district by weighing the global efficiency and redundancy of the network, and captures long-distance space Correlation eliminates the modeling over-smoothing problem caused by store similarity. The attention mechanism in the transformer is improved, and a new temporal feature dimension extractor is designed to resist the impact of time delays that may occur in the propagation of spatial information between locations. Using the multi-head graph attention mechanism, a new spatial feature dimension extractor is designed to dynamically capture the impact of different marketing methods and different commercial activities on traffic, and achieve dynamic coupling of spatial relationships between nodes.

Each embodiment in this specification is described in a related manner. The same and similar parts between the various embodiments can be referred to each other. Each embodiment focuses on its differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is simple. For relevant details, please refer to the partial description of the method embodiment.

The above descriptions are only preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Any modifications, substitutions, improvements, etc. made within the spirit and principles of the present invention are included in the protection scope of the present invention.

Claims (8)

1. A business district traffic prediction method based on six degrees of separation theory and graph attention network, which is characterized by including the following steps: S1. Collect business district traffic data; S2. Construct an adaptive business district traffic graph structure based on six degrees of separation theory; S3. Extract temporal features based on linear gated convolution attention unit; S4. Extract spatial features based on multi-head graph attention; S5. Output the prediction results through the fully connected layer to complete the construction of the business district traffic prediction model; S6. Business district traffic prediction model training; S7. Call the trained business district traffic prediction model to predict traffic. 2. The business district traffic prediction method based on six degrees of separation theory and graph attention network according to claim 1, characterized in that the business district traffic data collected by S1 includes business district range, shop location, collection frequency, collection period. 3. The business district traffic prediction method based on six degrees of separation theory and graph attention network according to claim 1, characterized in that the adaptive business district traffic graph structure construction process based on six degrees of separation theory in S2 is as follows: S21. Calculate the correlation coefficient of any two store node variables to obtain the affinity matrix S; normalize the affinity matrix S, then any element in the matrix The value is between 0 and 1; arrange the upper triangular matrix elements in the affinity matrix S in descending order; when When , change the corresponding position of the adjacency matrix A/> Set to 1, and the rest to 0, to obtain the adjacency matrix A;/> is the threshold; S22. Calculate the corresponding network average consistency for the adjacency matrix A. and the network average clustering coefficient CC , ultimately forming two curves; the intersection of the two curves/> A balance between efficiency and network redundancy;/> for/> The value of the network radius at the intersection point with the CC curve; S23. Set the balance point value as the optimal threshold, taking the distance to be greater than or equal to/> Node pairs are connected to form a business district traffic node graph structure. 4. The business district traffic prediction method based on six degrees of separation theory and graph attention network according to claim 1, characterized in that the S3 is based on the linear gated convolution attention unit to extract temporal features: first, transformer The self-attention mechanism in the model is replaced with local context-sensitive convolutional attention to form a gated attention unit to fit the time dimension characteristics of time series data; the input is a time length of Historical time series data, the input time series data is divided into blocks, and accurate convolution attention is used within the block and fast linear attention is used across blocks to obtain a gated attention unit with linear complexity; then by The degree-gated attention unit performs temporal convolution on the input data and outputs the feature vector of the correlation between each timestamp. 5. The business district traffic prediction method based on six degrees of separation theory and graph attention network according to claim 1, characterized in that the S4 extracting spatial features based on multi-head graph attention is specifically: Input the temporal feature vector obtained by S3 into the spatial convolution layer to calculate the feature vector with spatial characteristics. The specific calculation process is as follows: Among them,/> Is a matrix splicing operation; for the same node, multi-head self-attention is calculated separately/> attention and merged in a splicing or averaging manner attention;/> For the first/> Layer node/> eigenvector,/> For the first/> Layer node/> The feature vector of , after the aggregation operation with the attention mechanism as the core, the output is a node/> new eigenvector/> , understood as different nodes/> Characteristic Expression, Chapter/> nodes/> , the first one after going through the attention mechanism/> Node feature representation; Refers to the head of attention,/> Represents the attention coefficient,/> Represents node/> degree,/> ,/> is a learnable parameter,/> Is the first/> Vector representation of nodes, /> Indicates that the index is/> Middle/> Node pair/> The attention weight of the node;/> is the activation function,/> Is the first/> Vector representation of nodes; Attention coefficient The calculation method is as follows: Among them/> is a learnable parameter,/> No./> Attention coefficient of size,/> is an exponential function,/> Is a node/> eigenvector. 6. The business district traffic prediction method based on six degrees of separation theory and graph attention network according to claim 1, characterized in that the S5 specifically includes: merging temporal feature extraction and spatial feature vector through a fully connected layer and outputting the length. for Prediction results/> ;wherein,/> Represents the node index, /> Indicates time,/> Shown in/> Time node/> predicted value. 7. The business district traffic prediction method based on six degrees of separation theory and graph attention network according to claim 1, characterized in that the process of the S6 training model is specifically: S61 trains the model through the mean square error loss function. The loss function is: Among them/> is the total number of samples in the time series used to calculate the loss,/> is the true value,/> is the predicted value; S62. Use the Adam optimization algorithm to calculate the gradient of the network error for each weight parameter in backpropagation, and obtain new weights through the parameter update process; iteratively calculate the model weight until a predetermined small loss is reached, and the best prediction value is obtained. 8. The business district traffic prediction method based on six degrees of separation theory and graph attention network according to claim 1, characterized in that the S7 calls the trained business district traffic prediction model to perform traffic prediction: S71. Start real-time data collection, maintain a sampling frequency of once every 5 minutes, and input the collected data into the graph structure obtained by S2; S72. Call the model built in S6 to output the predicted traffic value.