Intelligent real-time advertisement pushing method based on multilevel graph neural network

Abstract

The invention belongs to the technical field of advertisement pushing, and relates to an intelligent real-time advertisement pushing method based on a multilevel graph neural network, which comprises the following steps of: firstly collecting dynamic data and static data of a parking lot and preprocessing, then designing two graph structures of a global graph and a local graph, and introducing a self-attention mechanism; a multi-level graph neural network is constructed and trained to obtain a trained model, finally, the trained model is utilized to calculate embedding of each license plate and a frequently-going place so as to generate a prediction score, for each user, advertisements are sorted according to the prediction score, and the advertisement with the highest score is selected as a push result. Through a multi-level graph structure, the model can capture a user behavior mode more comprehensively and meticulously, and the recommendation accuracy and correlation are improved; and moreover, the calculation complexity is reduced, the calculation efficiency of the model is improved, the method can be efficiently operated in an actual application scene with limited resources, and the requirement of real-time recommendation is met.

Description

Technical Field

[0001] The present invention belongs to the technical field of advertisement push, and relates to an intelligent real-time advertisement push method based on a multi-level graph neural network. Background Art

[0002] In modern parking lots, the traffic volume is large and the flow is frequent. When passing through the gate, car owners usually pay attention to the surrounding environment, especially the content on the gate display screen, which provides an excellent display opportunity for advertising. Since the vehicle stays for a short time, the advertisement must accurately attract the attention of the car owner in a short time to improve the conversion rate and effect of the advertisement. However, there is currently no algorithm on the market that specifically recommends advertisements based on the behavior of parking vehicles, resulting in low accuracy and relevance of advertising push, and the effect of advertising delivery has not been maximized.

[0003] Existing advertising recommendation algorithms are mostly used in e-commerce and social media scenarios, and cannot fully utilize parking lot-specific data resources, such as information about activity locations frequently visited by vehicles, vehicle brands and models, etc. These data resources are of great value in parking lot advertising recommendations because they can help the system understand the interests and needs of car owners more accurately. In existing technical solutions, although recommendation algorithms have made considerable progress, especially in the field of graph convolutional networks (GCN), they still have some significant shortcomings in practical applications: First, data integration is insufficient and data sources are single. Traditional recommendation systems mainly rely on a single type of data, such as users' historical ratings or browsing records. This method fails to fully utilize information from multiple data sources, such as dynamic information (frequently visited locations) and static information (vehicle brands and models) of vehicles, resulting in low accuracy and relevance of recommendations; second, the response speed is slow and the computational complexity is high. Existing graph neural network models, such as standard GCN, often contain complex feature transformations and nonlinear activation operations, which increase The computational complexity and training difficulty make it difficult to achieve fast response and instant recommendations in real-time application scenarios, especially in parking lots where vehicles stay for a very short time. In addition, due to computing resource and time constraints, traditional algorithms are prone to delays when processing large amounts of real-time data and cannot meet high real-time requirements. Third, the model complexity is high. Many existing recommendation systems use complex deep learning models that require a lot of computing resources and time for training and reasoning, which is a huge challenge for practical application scenarios with limited resources. In addition, the complex model structure makes it difficult to explain and debug, increasing the maintenance cost and difficulty of the model, and affecting the promotion of practical applications. In summary, traditional algorithms cannot effectively integrate and analyze this data, resulting in the content of advertising push being not relevant enough and unable to effectively attract the attention of car owners, thereby reducing the economic benefits of advertising. Summary of the invention

[0004] To solve the above problems, the present invention proposes an intelligent real-time advertising push method based on a multi-level graph neural network, which integrates the dynamic information and static information of the vehicle, and optimizes the information aggregation process through the self-attention mechanism to improve the accuracy and response speed of the recommendation, which is particularly suitable for real-time advertising push on parking gate display screens.

[0005] To achieve the above object, the present invention adopts the following technical solution: An intelligent real-time advertising push method based on a multi-level graph neural network specifically comprises the following steps: S1. Data collection: Collect dynamic and static data of the parking lot, where dynamic data includes vehicle entry and exit times and frequented activity locations, and static data includes vehicle brands and models; S2, data preprocessing: The dynamic data collected in step S1 is preprocessed using Apache Flink or Spark Streaming; S3. Multi-level graph neural network construction: Based on the CNN model, a global graph and a local graph are constructed respectively. The global graph is constructed based on the movement data of vehicles in different locations. The nodes include vehicles, locations and advertisements, and the edges represent the relationship between vehicles and locations, vehicles and advertisements, and locations and advertisements. The local graph includes the frequently visited location-advertising graph and the vehicle brand-advertising graph, which refines the advertising relationship between users' location and brand preferences. S4. Multi-level graph neural network training: Use historical data to pre-train the multi-level graph structure, adopt incremental update technology, and use online learning algorithms to update the embedding representation of users and items in real time to obtain a trained model; S5. Real-time advertising push: Use the trained model to calculate the embedding of each license plate and frequently visited places to generate a predicted score. For each user, sort the ads according to the predicted score and select the ads with the highest score as the push result.

[0006] As a further technical solution of the present invention, the dynamic data in step S1 is obtained through sensors and cameras in the parking lot, and the static data is obtained through a license plate recognition system.

[0007] As a further technical solution of the present invention, the global graph constructed in step S3 is expressed as , by the node set and edge sets Composition; Frequently visited places - advertising pictures (partial Figure 1 ) is expressed as , by the node set and edge sets Composition; Vehicle brand-advertising image (partial image 2) is represented as , by the node set and edge sets composition.

[0008] As a further technical solution of the present invention, the specific process of step S4 is: S41. Initialize the embedding representation of each node , , , , , Represents the global map, the frequently visited places-advertising map (local Figure 1 ) and the nodes in the vehicle brand-advertising graph (partial figure 2) The node embedding representation of layer 0 (initialization); S42. Perform graph convolution operations on the global graph to update the node embedding representations of advertisements, brands, and locations: , , , in , and They represent advertisement nodes, brand nodes, and location nodes respectively; , and Respectively represent nodes , and The set of neighbor nodes in the global graph; , and Respectively represent nodes , and In the global graph The node embedding representation of the layer; similarly, the graph convolution operation is performed on the two local graphs to obtain the updated node embedding representation: For frequently visited places - advertising pictures (partial Figure 1 ): , , For vehicle brand-advertising image (partial image 2): , , in and Respectively represent nodes and In the frequently visited places - advertising picture (partial Figure 1 ) in the neighbor node set; and Respectively represent nodes and The set of neighbor nodes in the vehicle brand-advertisement graph (partial figure 2); and Respectively represent nodes and In the frequently visited places - advertising picture (partial Figure 1 ) Node embedding representation of the layer; and Respectively represent nodes and In the vehicle brand-advertising picture (partial picture 2) Node embedding representation of the layer; S43, introduce the self-attention mechanism, and Figure 1 ) and the nodes in the vehicle brand-advertising graph (partial figure 2) Each layer of embedding representation , , Calculate the attention scores separately: , , ; in is the weight vector in the self-attention mechanism, is the weight matrix in the self-attention mechanism; S44, convert the attention score into attention weight through the Softmax function, so that the sum of all weights is 1, the global map, the frequently visited places-advertising map (local Figure 1 ) and vehicle brand-advertising image (partial image 2) The attention weights of the layer embedding representation are: , , ; S45. Weight the embedding representation of neighbor nodes with attention weights to form node The new embedding representation of is: , , ; S46, introduce multi-graph fusion mechanism, for each node , the node In the frequently visited places - advertising picture (partial Figure 1 ) and the vehicle brand-advertising image (partial figure 2) are fused through MLP to obtain the final embedding representation : ,in It is a learnable multi-layer perceptron; S47. Using Bayesian Personalized Ranking (BPR) loss function Calculation error: , in is the Sigmoid function, Indicates brand Frequently visited Location The prediction score of Indicates brand Frequently visited Location The prediction score of Controls the L2 regularization strength.

[0009] As a further technical solution of the present invention, the embedding of each license plate and frequently visited place calculated in step S5 is: , by calculating the inner product About Advertising Prediction score: .

[0010] Compared with the prior art, the present invention has the following advantages: (1) We designed two graph structures: a global graph and a local graph. The global graph captures the movement patterns of vehicles at different locations and their exposure to advertisements; the local graph focuses on specific details, such as the direct relationship between frequently visited locations and advertisements, and the association between vehicle brands and advertisements. This multi-level graph structure enables the model to capture user behavior patterns more comprehensively and meticulously, improving the accuracy and relevance of recommendations. It overcomes the problem that traditional GCN models usually only use a single graph structure and fail to fully integrate and analyze information from multiple data sources, resulting in less than ideal recommendation results.

[0011] (2) By introducing the self-attention mechanism, the model can adaptively assign weights according to the importance of nodes and edges, thereby capturing key information more effectively. The specific implementation of the self-attention mechanism includes the calculation of attention scores and the aggregation of information, so that the model can aggregate the information of neighbor nodes according to the calculated attention weights in each layer, and form the final embedding representation of each node through the layer combination strategy, which provides the accuracy and real-time performance of advertising recommendations and overcomes the problem that the traditional GCN model fails to fully utilize the attention mechanism in the process of information aggregation, resulting in weak information aggregation effect, which affects the accuracy and real-time performance of recommendations.

[0012] (3) The model architecture is simplified. By removing complex feature transformations and nonlinear activation operations, the computational complexity is reduced and the computational efficiency of the model is improved. It can run efficiently in actual application scenarios with limited resources and meet the needs of real-time recommendation. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 A flowchart of the intelligent real-time advertising push method based on a multi-level graph neural network provided by the present invention. DETAILED DESCRIPTION

[0014] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0015] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0016] It should be noted that the terms used herein are only for describing specific embodiments, and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that the terms "include" and "have" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0017] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0018] Example: like Figure 1 As shown, this embodiment provides an intelligent real-time advertisement push method based on a multi-level graph neural network, which specifically includes the following steps: S1. Data collection: Collect dynamic data and static data of the parking lot. Dynamic data includes the entry and exit time of vehicles and the places where they frequently go, which are obtained through sensors and cameras in the parking lot. Static data includes the brand and model of vehicles, which are obtained through the license plate recognition system.

[0019] S2. Data preprocessing: The dynamic data collected in step S1 is preprocessed using Apache Flink or Spark Streaming.

[0020] S3. Multi-level graph neural network construction: Based on the CNN model, we construct global graph and local graph respectively. It is constructed based on the movement data of vehicles at different locations. Nodes include vehicles, locations, and advertisements. Edges represent the relationship between vehicles and locations, vehicles and advertisements, and locations and advertisements. All nodes constitute a node set. , all edges form an edge set ; Partial map includes frequently visited places - advertising map (partial Figure 1 ) and vehicle brand-advertising image (partial image 2) , refine the advertising relationship between users in terms of location and brand preferences, Frequently visited places - advertising picture (partial Figure 1 )’s node set and edge set; are the node sets and edge sets of the vehicle brand-advertisement graph (partial figure 2).

[0021] S4. Multi-level graph neural network training: Use historical data to pre-train the multi-level graph structure, adopt incremental update technology, and use online learning algorithms to update the embedding representation of users and items in real time to obtain a trained model; the specific process is as follows: S41. Initialize the embedding representation of each node , , , , , Represents the global map, the frequently visited places-advertising map (local Figure 1 ) and the nodes in the vehicle brand-advertising graph (partial figure 2) The node embedding representation of layer 0 (initialization); S42. Perform graph convolution operations on the global graph to update the node embedding representations of advertisements, brands, and locations: , , , in , and They represent advertisement nodes, brand nodes, and location nodes respectively; , and Respectively represent nodes , and The set of neighbor nodes in the global graph; , and Respectively represent nodes , and In the global graph The node embedding representation of the layer; similarly, the graph convolution operation is performed on the two local graphs to obtain the updated node embedding representation: For frequently visited places - advertising pictures (partial Figure 1 ): , , For vehicle brand-advertising image (partial image 2): , , in and Respectively represent nodes and In the frequently visited places - advertising picture (partial Figure 1 ) in the neighbor node set; and Respectively represent nodes and The set of neighbor nodes in the vehicle brand-advertisement graph (partial figure 2); and Respectively represent nodes and In the frequently visited places - advertising picture (partial Figure 1 ) Node embedding representation of the layer; and Respectively represent nodes and In the vehicle brand-advertising picture (partial picture 2) Node embedding representation of the layer; S43. In the global graph and the local graph, the relationship and importance between different nodes are different. The traditional graph convolution operation assigns the same weight to all neighbor nodes, ignoring the importance differences of different neighbor nodes. The attention mechanism can adaptively assign weights according to the characteristics of the nodes and assign different weights to each neighbor node, thereby more accurately reflecting its importance in information aggregation. This can effectively process heterogeneous data and enable the model to better adapt to different types of graph structures and relationships. Therefore, this embodiment introduces the self-attention mechanism to the global graph, the frequently visited places-advertising graph (local Figure 1 ) and the nodes in the vehicle brand-advertising graph (partial figure 2) Each layer of embedding representation , , Calculate the attention scores separately: , , ; in is the weight vector in the self-attention mechanism, is the weight matrix in the self-attention mechanism; S44, convert the attention score into attention weight through the Softmax function, so that the sum of all weights is 1, the global map, the frequently visited places-advertising map (local Figure 1 ) and vehicle brand-advertising image (partial image 2) The attention weights of the layer embedding representation are: , , ; S45. Weight the embedding representation of neighbor nodes with attention weights to form node The new embedding representation of is: , , ; S46. In order to unify the node embedding representation of the global graph and the local graph, a multi-graph fusion mechanism is introduced. , the node In the frequently visited places - advertising picture (partial Figure 1 ) and the vehicle brand-advertising image (partial figure 2) are fused through MLP to obtain the final embedding representation : ,in It is a learnable multi-layer perceptron; S47. Using Bayesian Personalized Ranking (BPR) loss function Computational Error, Bayesian Personalized Ranking (BPR) Loss Function It is a pairwise loss function that encourages observed interaction terms to have higher prediction scores than unobserved interaction terms. Its formula is: , in is the Sigmoid function, Indicates brand Frequently visited Location The prediction score of Indicates brand Frequently visited Location The prediction score of Controls the L2 regularization strength.

[0022] S5, real-time advertising push: use the trained model to calculate the embedding of each license plate and frequently visited location: , by calculating the inner product About Advertising Prediction score: ,For each user, ads are sorted according to the predicted ratings, and the ads with the highest ratings are selected as push results.

[0023] This embodiment combines the dynamic data and static data of the vehicle for the first time, integrates the currently available data, and utilizes the dynamic information of the vehicle, such as the vehicle's frequent activity locations, and static information, such as the prior knowledge of the vehicle brand. By combining the multi-level graph structure and the self-attention mechanism, the improved lightweight graph convolutional network method can make more accurate advertising recommendations, which is particularly suitable for real-time advertising push on parking lot gate display screens. This method not only improves the accuracy of the recommendation, but also maintains a high computational efficiency. It is suitable for application scenarios with high real-time and high precision requirements. Through the multi-level graph structure, the complex relationship between vehicles and advertisements can be captured more comprehensively. The global graph provides an overall perspective, capturing the movement patterns and advertising exposure of vehicle brands in different locations; the local graph focuses on specific details, such as the direct relationship between frequently visited locations and advertisements, and the association between vehicle brands and advertisements. This multi-level graph structure enables the model to capture user behavior patterns more meticulously and improve the accuracy and relevance of recommendations; and a self-attention mechanism is designed for the heterogeneity of the global graph and the local graph. The introduction of the self-attention mechanism further optimizes the information aggregation process. Through the self-attention mechanism, the model can adaptively assign weights according to the importance of nodes and edges, so as to capture key information more effectively. The specific implementation of the self-attention mechanism includes the calculation of attention scores and the aggregation of information, so that the model can aggregate the information of neighbor nodes according to the calculated attention weights in each layer, and form the final embedding representation of each node through the layer combination strategy, which solves the problem of real-time advertising push on the parking lot gate display screen. It can not only make full use of the advantage of high traffic volume in the parking lot, but also improve the relevance and economic benefits of advertising through precise recommendation algorithms, which not only brings higher returns to advertisers, but also improves the user experience of car owners. It has important practical application value.

[0024] Those skilled in the art will appreciate that the units, i.e., algorithm steps, of the various examples described in conjunction with this embodiment can be implemented in electronic hardware or in a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0025] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

[0026] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. An intelligent real-time advertising push method based on a multi-level graph neural network, characterized in that: The specific steps include: S1. Data collection: Collect dynamic and static data of the parking lot, where dynamic data includes vehicle entry and exit times and frequented activity locations, and static data includes vehicle brands and models; S2, data preprocessing: The dynamic data collected in step S1 is preprocessed using Apache Flink or Spark Streaming; S3. Construction of multi-level graph neural network: Based on the CNN model, a global graph and a local graph are constructed respectively. The global graph is constructed based on the movement data of vehicles in different locations. The nodes include vehicles, locations and advertisements, and the edges represent the relationship between vehicles and locations, vehicles and advertisements, and locations and advertisements. The local graph includes the frequently visited location-advertising graph and the vehicle brand-advertising graph, which refines the advertising relationship between users' location and brand preferences. S4. Multi-level graph neural network training: Use historical data to pre-train the multi-level graph structure, adopt incremental update technology, and use online learning algorithms to update the embedding representation of users and items in real time to obtain a trained model; S5. Real-time advertising push: Use the trained model to calculate the embedding of each license plate and frequently visited places to generate a predicted score. For each user, sort the ads according to the predicted score and select the ads with the highest score as the push result.

2. The intelligent real-time advertising push method based on a multi-level graph neural network according to claim 1 is characterized in that: The dynamic data in step S1 is obtained through sensors and cameras in the parking lot, and the static data is obtained through a license plate recognition system.

3. The intelligent real-time advertising push method based on a multi-level graph neural network according to claim 2 is characterized in that: The global graph constructed in step S3 is represented as , by the node set and edge sets Composition; Frequently visited places - advertising map is represented as , by the node set and edge sets Composition; Vehicle brand - advertising image is represented by , by the node set and edge sets composition.

4. The intelligent real-time advertising push method based on a multi-level graph neural network according to claim 3 is characterized in that: The specific process of step S4 is: S41. Initialize the embedding representation of each node , , , , , Represent nodes in the global graph, the frequently visited places-advertising graph, and the vehicle brand-advertising graph respectively The node embedding representation of the 0th layer; S42. Perform graph convolution operations on the global graph to update the node embedding representations of advertisements, brands, and locations: ， ， ， in , and They represent advertisement nodes, brand nodes, and location nodes respectively; , and Respectively represent nodes , and The set of neighbor nodes in the global graph; , and Respectively represent nodes , and In the global graph The node embedding representation of the layer; similarly, the graph convolution operation is performed on the two local graphs to obtain the updated node embedding representation: For frequently visited places - advertising pictures: ， ， For vehicle brand-advertising image: ， ， in and Respectively represent nodes and The set of neighbor nodes in the frequented places-advertisement graph; and Respectively represent nodes and The set of neighbor nodes in the vehicle brand-advertisement graph; and Respectively represent nodes and In the frequently visited places - the first Node embedding representation of the layer; and Respectively represent nodes and In the vehicle brand-advertising picture Node embedding representation of the layer; S43, introduce the self-attention mechanism to the nodes in the global graph, the frequently visited places-advertising graph, and the vehicle brand-advertising graph Each layer of embedding representation , , Calculate the attention scores separately: ， ， ； in is the weight vector in the self-attention mechanism, is the weight matrix in the self-attention mechanism; S44. The attention scores are converted into attention weights through the Softmax function so that the sum of all weights is 1. The attention weights of the layer embedding representation are: ， ， ； S45. Weight the embedding representation of neighbor nodes with attention weights to form node The new embedding representation of is: ， ， ； S46, introduce multi-graph fusion mechanism, for each node , the node The embedding representations of the frequently visited places-advertising images and vehicle brands-advertising images are fused through MLP to obtain the final embedding representation. : ,in It is a learnable multi-layer perceptron; S47. Use Bayesian personalized ranking loss function Calculation error: ， in is the Sigmoid function, Indicates brand Frequently visited Location The prediction score of Indicates brand Frequently visited Location The prediction score of Controls the L2 regularization strength.

5. The intelligent real-time advertising push method based on a multi-level graph neural network according to claim 4 is characterized in that: The embedding of each license plate and frequently visited place calculated in step S5 is: , by calculating the inner product About Advertising Prediction score: .

Images