Advertisement recommendation method and system based on fusion neural network

Abstract

The present application provides an advertising recommendation method and system based on a fusion neural network. Among them, the real-time behavior data of the user is received; based on the real-time behavior data, a user interest map is dynamically constructed, and the user interest map includes the user's nodes of interest and the associations between the nodes of interest; the user interest map is analyzed in real time using a fusion neural network model; based on the preference probability, a set of highly relevant advertisements is screened out as candidate advertisements; combined with the current scenario of the user, a target advertisement is selected from the candidate advertisements for push, so as to improve the click-through rate and conversion rate of the target advertisement. The technical solution provided by the present application can effectively improve the recommendation effect of advertisements.

Description

Technical Field

[0001] The embodiments of the present application relate to the technical field of advertisement recommendation, and in particular, to an advertisement recommendation method and system based on a fusion neural network. Background Art

[0002] With the popularization of the Internet and mobile devices, digital marketing has become an important means for companies to promote their products and services.

[0003] The traditional recommendation algorithms based on historical data are commonly used in the market, such as collaborative filtering, content-based recommendation, etc. These methods usually collect users' browsing history, purchase history and other interactive behaviors, and then calculate the similarity or matching degree through algorithms to recommend products or services that may be of interest to users.

[0004] However, traditional recommendation algorithms mainly rely on historical behavior data of users, lack sensitivity to real-time behavior, and cannot timely reflect the current changes in users' interests. In addition, these algorithms often ignore the impact of different scenarios of users on recommendation results, resulting in the recommended ads not meeting the current needs of users, reducing the click-through rate and conversion rate of ads. Summary of the invention

[0005] The embodiments of the present application provide an advertisement recommendation method and system based on a fusion neural network, so as to solve the problem of poor advertisement recommendation effect in the prior art.

[0006] In a first aspect, the present application provides an advertisement recommendation method based on a fusion neural network, comprising:

[0007] Receive real-time behavior data of users; based on the real-time behavior data, dynamically construct a user interest graph, wherein the user interest graph includes nodes of interest to the user and associations between nodes of interest; use a fusion neural network model to analyze the user interest graph in real time, wherein the fusion neural network model integrates the user's historical behavior data and current scene information to predict the user's preference probability for a specific advertising category; based on the preference probability, screen out a set of highly relevant advertisements as candidate advertisements; in combination with the current scene of the user, select a target advertisement from the candidate advertisements for push, so as to improve the click-through rate and conversion rate of the target advertisement; wherein the fusion neural network model is also used to optimize the advertisement recommendation strategy according to the feedback mechanism to ensure that the recommended target advertisement is highly matched with the user's interests.

[0008] In a second aspect, the embodiment of the present application provides an advertisement recommendation system based on a fusion neural network, comprising:

[0009] A receiving module, used to receive real-time behavior data of users;

[0010] A construction module, used to dynamically construct a user interest graph based on the real-time behavior data, wherein the user interest graph includes nodes of interest to the user and associations between nodes of interest;

[0011] An analysis module, used to analyze the user interest graph in real time using a fusion neural network model, wherein the fusion neural network model is used to predict the user's preference probability for different advertising categories by integrating the user's historical behavior data and current scene information;

[0012] A screening module, used to screen out a set of highly relevant advertisements as candidate advertisements based on the preference probability;

[0013] The push module is used to select a target advertisement from the candidate advertisements for push based on the current scenario of the user, so as to improve the click-through rate and conversion rate of the target advertisement; wherein the fusion neural network model is also used to optimize the advertisement recommendation strategy according to the feedback mechanism to ensure that the recommended target advertisement is highly matched with the user's interests.

[0014] In a third aspect, an embodiment of the present application provides a computing device, comprising a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement an advertising recommendation method based on a fusion neural network as described in the first aspect.

[0015] In a fourth aspect, an embodiment of the present application provides a computer storage medium storing a computer program. When the computer program is executed by a computer, it implements an advertisement recommendation method based on a fusion neural network as described in the first aspect.

[0016] In the embodiment of the present application, the user's real-time behavior data is received; based on the real-time behavior data, a user interest map is dynamically constructed, the user interest map includes the user's nodes of interest and the associations between the nodes of interest; the user interest map is analyzed in real time using a fusion neural network model; based on the preference probability, a highly relevant set of advertisements is screened out as candidate advertisements; combined with the current scenario of the user, a target advertisement is selected from the candidate advertisements for push, so as to improve the click-through rate and conversion rate of the target advertisement. The technical solution provided by the present application can effectively improve the recommendation effect of advertisements.

[0017] These and other aspects of the present application will become more clearly understood in the description of the following embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0019] Figure 1 A flowchart of an advertisement recommendation method based on a fusion neural network provided in an embodiment of the present application;

[0020] Figure 2 A schematic diagram of the structure of an advertisement recommendation system based on a fusion neural network provided in an embodiment of the present application;

[0021] Figure 3 A schematic diagram of the structure of a computing device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0022] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0023] In some of the processes described in the specification and claims of this application and the above-mentioned figures, multiple operations that appear in a specific order are included, but it should be clearly understood that these operations may not be executed in the order in which they appear in this article or executed in parallel. The serial numbers of the operations, such as 101, 102, etc., are only used to distinguish between different operations, and the serial numbers themselves do not represent any execution order. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions of "first", "second", etc. in this article are used to distinguish different messages, devices, modules, etc., do not represent the order of precedence, and do not limit the "first" and "second" to be different types.

[0024] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of this application.

[0025] Figure 1 A flowchart of an advertisement recommendation method based on a fusion neural network is provided for an embodiment of the present application. Figure 1 As shown, the method includes:

[0026] 101. Receive real-time behavior data of users;

[0027] In this step, real-time behavior data refers to the instant behavior records generated by users when using applications, websites or other digital platforms, including but not limited to clicks, browsing, searches, purchases, comments, etc. This data is usually collected through front-end embedding technology and can reflect the user's current interests and behavior patterns.

[0028] In the embodiment of the present application, firstly, each interactive behavior of the user, such as page clicks, swipes, dwell time, etc., is captured in real time through the front-end JavaScript code or SDK. Secondly, the collected real-time behavior data is transmitted to the back-end server through an API or a message queue (such as Kafka). Finally, the received data is stored in a real-time data warehouse (such as ClickHouse, Apache Doris) for subsequent processing and analysis.

[0029] 102. Based on the real-time behavior data, dynamically construct a user interest graph, wherein the user interest graph includes nodes of interest to the user and associations between nodes of interest;

[0030] In this step, the user interest graph is a graph structure, in which nodes represent user interests (such as products, articles, videos, etc.), and edges represent the associations between these interests. The graph can more intuitively display the user's interest preferences and behavior paths. Dynamic construction means updating and adjusting the user interest graph in real time based on the user's real-time behavior data to reflect the user's latest interest changes.

[0031] In the embodiment of the present application, first, the real-time behavior data is cleaned and preprocessed to extract key user behavior information. Secondly, based on the user's clicks, browsing and other behaviors, the user's points of interest are identified and used as nodes in the graph. Furthermore, by analyzing the user's behavior path between different points of interest, the association relationship between nodes is established to form a graph. Finally, as the user's behavior continues to change, the nodes and edges in the graph are updated in real time to ensure the timeliness and accuracy of the graph.

[0032] The embodiment of the present invention proposes a further optional solution, which is to extract behavioral features of more dimensions by deeply mining real-time behavioral data, and to form a more detailed and comprehensive user interest network by building correlations between behavioral features.

[0033] The options are as follows:

[0034] Optionally, the step 102 of dynamically constructing a user interest graph based on the real-time behavior data includes: extracting user behavior characteristics from the real-time behavior data, the behavior characteristics including the user's browsing habits, purchasing tendencies, and social interactions; mapping the behavior characteristics to nodes in the user interest graph, and establishing connecting edges between the nodes based on similarity or co-occurrence frequency to construct the user interest graph.

[0035] Suppose there is a user's real-time behavior data, which includes the user's online activity records in the past week. The following behavioral features can be extracted from this data: the user frequently visits sports goods websites; recently purchased several sports equipment; follows multiple sports stars on social media, and often comments and shares sports event videos. Next, these behavioral features are mapped to nodes in the user's interest graph: node A is sports goods; node B is sports equipment; node C is sports stars; node D is sports events. Then, based on the similarity and co-occurrence frequency between user behavior features, connect edges between nodes: edge AB is the user's high frequency of purchasing sports equipment, indicating that there is a strong association between the two nodes; edge AC is the user following sports stars and interacting with them, indicating that the two nodes also have a strong relationship; edge AD is the user sharing and commenting on sports event videos, which also reflects the connection between the two nodes.

[0036] Through the above steps, the system can more finely capture the multi-dimensional behavioral characteristics of users, and form a more detailed and comprehensive user interest network by building the correlation between behavioral characteristics. This not only improves the accuracy of recommendation results, but also enhances user experience and solves the problem of inaccurate recommendation results in traditional methods.

[0037] 103. Analyze the user interest graph in real time using a fusion neural network model, wherein the fusion neural network model is used to predict the user's preference probability for different advertising categories by integrating the user's historical behavior data and current scene information;

[0038] In this step, the fusion neural network model is a model that combines multiple neural network technologies to integrate the user's historical behavior data and current scene information to predict the user's preference probability for different advertising categories. Historical behavior data refers to the user's behavior records in the past period of time, including clicks, purchases, comments, etc. Current scene information refers to the user's current context information, such as time, location, device type, etc.

[0039] In the embodiment of the present application, first, the user's historical behavior data and current scene information are integrated together to form an input feature vector. Secondly, the nodes and edges in the user's interest graph are converted into vector representations, and the features of the graph are extracted through a graph neural network (GNN). Furthermore, the historical behavior data, current scene information, and graph features are fused through a multi-layer neural network to generate a comprehensive feature vector. Finally, through activation functions such as softmax, the comprehensive feature vector is converted into the user's preference probability for different advertising categories.

[0040] The embodiment of the present invention proposes an optional solution, which integrates the user's real-time behavior data, historical behavior data and current scene information through a fusion neural network model to more accurately predict the user's preference probability for different types of advertisements, thereby improving the relevance and attractiveness of recommended advertisements.

[0041] The options are as follows:

[0042] Optionally, the use of a fusion neural network model to analyze the user interest map in real time in step 103 includes: inputting the user's real-time behavior data and the acquired historical behavior data into a pre-trained fusion neural network model, so as to capture the non-linear relationship in the user's behavior and the interdependence between different behaviors through the fusion neural network model using a multi-level hidden layer; based on the non-linear relationship and the interdependence, and in combination with current scene information, determining the user's preference probability for different types of advertisements, wherein the current scene information is used to increase the model's ability to predict user behavior in specific scenarios.

[0043] Assume that user Xiao Li has recently frequently visited the webpage of fitness equipment and shared some fitness-related posts on social media. At the same time, Xiao Li's historical behavior data shows that he has purchased running shoes and sportswear. Today is the weekend, the weather is fine, and Xiao Li is taking a walk in the park.

[0044] Input real-time behavior data and historical behavior data: Xiao Li viewed fitness equipment today and shared fitness posts; Xiao Li has purchased running shoes and sportswear and browsed fitness equipment; the model captures Xiao Li's behavioral characteristics through multiple hidden layers, such as interest in fitness and inclination to purchase sports equipment; the model also captures the nonlinear relationship and interdependence between Xiao Li's behaviors, such as being more likely to share fitness posts after viewing fitness equipment; the current scene information shows that Xiao Li is walking in the park and the weather is sunny. Combining this information, the model predicts that Xiao Li has a high probability of preferring outdoor sports equipment and fitness equipment advertisements.

[0045] By integrating the user's real-time behavior data, historical behavior data, and current scene information into a fusion neural network model, the system can more accurately capture the nonlinear relationships and interdependencies in user behavior, thereby more accurately predicting the probability of user preference for different types of ads. This not only improves the relevance and attractiveness of recommended ads, but also enhances the user experience and solves the problem of inaccurate recommendation results in traditional recommendation methods.

[0046] The present application takes into account that although the existing advertising recommendation system can predict the user's advertising preferences based on the user's historical behavior data, it is insufficient in terms of real-time and contextual awareness. Specifically, existing systems often find it difficult to capture changes in user behavior in real time, and often ignore the user's current specific context when recommending advertisements. This results in low accuracy of recommendation results and poor user experience. Therefore, an embodiment of the present invention proposes an optional solution, which introduces a fusion neural network model and combines real-time behavior data, historical behavior data, and current scene information to predict the user's preference probability for a specific advertising category, thereby improving the relevance and attractiveness of recommended advertisements.

[0047] The options are as follows:

[0048] Optionally, the process of predicting the user's preference probability for a specific advertisement category by the fusion neural network model in step 103 includes:

[0049] The probability of the user's preference for a specific advertisement category is calculated by the following formula:

[0050] ;

[0051] in, Indicates that the user is at time For specific ad categories The probability of preference; It is the user's real-time behavior data and historical behavior data in time. The eigenvector of For specific ad categories The weight matrix of For specific ad categories The bias term of Is the user at time Current scene information, including geographic location, time, weather conditions, and device status; It is the weight matrix used to fuse the current scene information; ReLU function , used to fuse the user's feature vector and current scene information ; is the regularization coefficient, used to control the weight matrix The size of , to prevent overfitting; is the weight matrix The L2 norm squared; is the set of all advertising categories; For specific ad categories The bias, is a weight matrix for a specific ad category; as well as In Represents the transpose of the weight matrix.

[0052] The overall design goal of this formula is to accurately predict the probability of users' preference for specific advertising categories by integrating users' real-time and historical behavior data, current scenario information, and combining neural network models. This helps online advertising platforms better understand user needs, thereby increasing the click-through rate and conversion rate of advertisements, and improving user experience and commercial value.

[0053] The following is a brief introduction to the design reasons of each item of the formula:

[0054] exist middle, Used to calculate the user's feature vector For specific ad categories Contribution Used to calculate the current scene information For specific ad categories Contribution Baseline used to adjust overall ratings; Used to prevent overfitting.

[0055] The following is a brief introduction to how to obtain the parameters of the formula:

[0056] User feature vector : It is obtained by collecting historical behavior data such as users' browsing history, search history, click history, etc., combining it with users' real-time behavior data and processing it through feature engineering.

[0057] Weight Matrix and bias : It is obtained by training a neural network model. The input of the model is the user feature vector and current scene information, and the output is the user's preference probability for each advertising category.

[0058] Current scene information : Obtain the user's geographic location, time, weather conditions and other information in real time through the API interface.

[0059] Weight Matrix : It is also obtained by training the neural network model and is used to fuse the current scene information.

[0060] Regularization coefficient : Determine the optimal value through cross-validation and other techniques to balance the complexity and generalization ability of the model.

[0061] Ad Category Collection :Determined according to the classification standards of the advertising platform, each advertising category corresponds to a unique identifier.

[0062] In the embodiment of the present application, it is assumed that the above formula is implemented on an online advertising platform, and the specific parameters are as follows:

[0063] User feature vector (Simplified representation)

[0064] Current scene information (e.g., Location: City Center, Time: Afternoon, Weather: Sunny, Device: Mobile)

[0065] Ad Category Collection

[0066] Weight Matrix and and bias Obtained by training the model, assuming the following:

[0067]

[0068] Regularization coefficient

[0069] First, calculate the unnormalized preference score for each ad category:

[0070]

[0071] Among them, the specific calculation of the unnormalized preference score of the advertising category (entertainment) is as follows:

[0072]

[0073] Similarly, calculate the unnormalized preference scores for other ad categories:

[0074]

[0075] Normalization:

[0076]

[0077] Calculation results:

[0078]

[0079] According to the calculation results, users have the highest probability of preferring "technology" ads in the current scenario (about 0.3010), followed by "fashion" (about 0.2462), "entertainment" (about 0.2298) and "news" (about 0.2230). This shows that in the current scenario, users are most likely to be interested in technology ads, so this type of ads can be displayed first to increase click-through rate and conversion rate.

[0080] 104. Screening out a set of highly relevant advertisements as candidate advertisements based on the preference probability;

[0081] In this step, the preference probability refers to the user's preference probability for each advertisement category calculated by the fusion neural network model. The highly relevant advertisement set refers to selecting advertisements in advertisement categories with higher preference probabilities according to the preference probability to form a candidate advertisement set.

[0082] In the embodiment of the present application, first, a threshold value of preference probability (e.g., 0.2) is set according to business needs to filter out highly relevant advertising categories. Secondly, advertising categories with preference probabilities higher than the threshold are filtered out. Finally, specific advertisements are selected from the filtered advertising categories to be added to the candidate advertising set. The top N advertisements in each advertising category can be selected (e.g., the top 5 advertisements in each category).

[0083] Furthermore, the embodiment of the present invention proposes an optional solution, which ensures that the selected candidate advertisements are highly relevant to the user's preferences through clear screening steps and standards, thereby improving the quality of recommended advertisements.

[0084] The options are as follows:

[0085] Optionally, the step 104 of filtering out a set of highly relevant advertisements as candidate advertisements based on the preference probability includes: sorting the preference probabilities of different types of advertisements to filter out advertisement categories with preference probabilities higher than a preset threshold; constructing an advertisement library, the advertisement library containing all optional advertisements and information corresponding to the optional advertisements, the information corresponding to the optional advertisements including advertisement category labels, advertisement content descriptions, and target audience group definitions; matching the filtered advertisement categories with the optional advertisements in the advertisement library based on a pre-constructed advertisement library index to find an advertisement set that matches the user preference, wherein the advertisement library index is used to match the optional advertisements in the advertisement library with the user preference based on the advertisement category label, content description, and target audience group definition; sorting the matched advertisement set according to the comprehensive score of the advertisement relevance and the user preference, and selecting advertisements with scores higher than a preset score as candidate advertisements.

[0086] Assume that user Xiao Zhang has recently frequently browsed information about digital products and discussed laptops on social media. Based on real-time behavior data and historical behavior data, the system calculates that Xiao Zhang's preference probability for digital product ads is 0.85, while the preset threshold is 0.5; Xiao Zhang's preference probability for digital product ads is 0.85, which is much higher than the preset threshold of 0.5, so the digital product ad category is screened out; the ad library contains multiple ads, each of which has detailed description information, such as ad category labels, ad content descriptions, target audience group definitions, etc.; in the ad library, find all ads related to digital products, which constitute a preliminary ad set; suppose there are three ads related to digital products in the ad library, with comprehensive scores of 90, 85, and 70 respectively; the preset score is set to 80, so ads with scores higher than 80 (90 and 85) are selected as candidate ads.

[0087] Through the above steps, the system can more finely filter out advertisements that are highly relevant to user preferences as candidate advertisements. This method not only improves the relevance and attractiveness of recommended advertisements, but also enhances user experience and solves the problem of inaccurate recommendation results in traditional recommendation methods. Ultimately, the advertisements that users see will be more in line with their interests, thereby increasing the click-through rate and conversion rate of advertisements.

[0088] 105. In combination with the current scenario of the user, a target advertisement is selected from the candidate advertisements for push, so as to improve the click-through rate and conversion rate of the target advertisement; wherein the fusion neural network model is also used to optimize the advertisement recommendation strategy according to the feedback mechanism, so as to ensure that the recommended target advertisement is highly matched with the user's interests.

[0089] In this step, the current scene information refers to the user's geographic location, time, weather conditions, device status and other information, which can further refine the advertising recommendation strategy. Targeted advertising refers to selecting the most suitable advertisement for the current scene from the candidate advertisement set for push. The feedback mechanism refers to continuously optimizing the advertising recommendation strategy by collecting user click and conversion data to ensure that the recommended target advertisement is highly matched with the user's interests.

[0090] In an embodiment of the present application, first, the user's geographic location, time, weather conditions, device status and other information are obtained in real time through an API interface. Secondly, based on the current scene information, the relevance score of each candidate advertisement with the current scene is calculated. A pre-trained model or rule can be used to calculate the relevance score. Furthermore, from the candidate advertisement set, the advertisement with the highest scene relevance score is selected as the target advertisement for push. The user's click and conversion data are recorded as feedback data. Finally, based on the feedback data, the model parameters are adjusted or the recommendation algorithm is optimized to gradually improve the click-through rate and conversion rate of the advertisement.

[0091] The embodiment of the present invention proposes an optional solution, which optimizes the advertisement recommendation strategy by combining the user's current scenario information to improve the click-through rate and conversion rate of the advertisement.

[0092] The options are as follows:

[0093] Optionally, the method in step 105 combines the current scene in which the user is located to select a target advertisement from the candidate advertisements for push so as to improve the click-through rate and conversion rate of the target advertisement, including: obtaining scene information corresponding to the current scene in which the user is located through an environmental perception module, including geographic location, time, weather conditions, and device status; analyzing the scene information, extracting key scene features, and combining the key scene features with the real-time behavior data to form comprehensive context data; inputting the comprehensive context data into the fusion neural network model to evaluate the expected performance of the candidate advertisements in the current scene to obtain an evaluation result; based on the evaluation result, selecting a target advertisement from the candidate advertisements that best matches the current scene and has the highest potential click-through rate and conversion rate; formulating a corresponding advertisement push strategy based on the target advertisement, the advertisement push strategy including a push time window, a push form, and a push channel; and executing the advertisement push strategy to improve the click-through rate and conversion rate of the target advertisement.

[0094] Suppose user Xiao Zhao is surfing the Internet on his mobile phone in a cafe near his home at around 8 pm on a Friday. Based on real-time behavioral data, the system knows that Xiao Zhao has recently frequently browsed information about movies and discussed upcoming movies on social media. At this time, the system obtains Xiao Zhao's scene information as follows: the geographical location is a cafe in the city center; the time is 8 pm on Friday; the weather condition is sunny; the device status is using a mobile phone to surf the Internet; the system obtains Xiao Zhao's current geographical location, time, weather conditions, and device status; extracts key scene features, namely Friday night, being in a cafe, and using a mobile phone; combines the key scene features with Xiao Zhao's real-time behavioral data (such as recently browsed movie information) to form comprehensive context data; inputs the comprehensive context data into the fusion neural network model to evaluate the expected performance of candidate ads; assumes that the evaluation results show that the local cinema's ads have the best expected performance in the current scenario; selects the local cinema's ads as the target ads; pushes ads through social media at 8:30 pm; pushes ads through specified channels at specified times to increase the click-through rate and conversion rate of ads.

[0095] By combining the user's current scenario information to optimize the ad recommendation strategy, the system can more accurately select the target ad that best suits the current scenario and formulate an effective push strategy, thereby significantly improving the click-through rate and conversion rate of the ad. This method not only enhances the user experience, but also improves the effectiveness of ad delivery, solving the problem of inaccurate recommendation results in traditional recommendation methods.

[0096] The present application takes into account that existing advertising recommendation systems often rely only on historical behavior data of users when recommending advertisements, but fail to fully utilize the specific scenario information of the user's current location. Since users have different needs and behavior patterns in different scenarios, relying solely on historical data is not enough to fully and accurately predict the user's current interests and needs. In order to solve this problem, an embodiment of the present invention proposes an optional solution to optimize the advertising recommendation strategy by combining the user's current scenario information to improve the click-through rate and conversion rate of the advertisement.

[0097] The options are as follows:

[0098] Optionally, the fusion neural network model evaluates the candidate advertisement in the current scenario and in expected performance, and the process of obtaining the evaluation result includes:

[0099] The candidate advertisement's expected performance in the current scenario is calculated using the following formula to obtain an evaluation result:

[0100] ;

[0101] in, Indicates at time and comprehensive contextual data Candidate ads under the assessment results; Is a candidate ad The embedding vector of ; is a nonlinear function used to integrate the user feature vector and current situation information , generating a comprehensive vector; is a series of indicator functions, each indicator function corresponds to a candidate ad In comprehensive contextual data An attribute in is each indicator function The weight factor is used to adjust the importance of attributes corresponding to different comprehensive context data; Is a set of functions used to calculate user feature vectors Scene information related to current situation information The strength of the interaction between Each function The weight factor is used to adjust the importance of the interaction between different features; It is a regularization term used to penalize the complexity of the user feature vector and the current context information to prevent overfitting; is the regularization term The coefficient of

[0102] in, , Represents the user feature vector and current situation information Connect into a vector; and is the weight matrix; and is the bias vector; is the activation function.

[0103] in, property With advertising In context Next match , It is a candidate ad; It is comprehensive contextual data; is the index of the indicator function, corresponding to an attribute of the advertisement; Is an indicator function that returns 1 if the condition is met, otherwise it returns 0.

[0104] in,

[0105] , is the weight matrix used to calculate the user feature vector With scene information Interaction strength; is the weight matrix used to calculate scene information and user feature vector Interaction strength; is a bias term used to adjust the interaction strength; It is an activation function, usually a nonlinear function such as ReLU, sigmoid or tanh.

[0106] In the ad recommendation system, evaluating the expected performance of candidate ads in the current scenario is a key step to improve ad click-through rate and conversion rate. This formula aims to evaluate the performance of candidate ads by integrating the neural network model, taking into account user characteristics, current scenario information, ad attributes, and user-scenario interaction.

[0107] The reason for the overall design of the formula is to comprehensively evaluate the performance of candidate ads by taking into account user characteristics, current scenario information, ad attributes, and user-scenario interactions. Capture the complex relationship between user characteristics and current scene information. and Adjust the importance of different attributes and interactions. Introduce regularization terms Prevent model overfitting and ensure that the model has good generalization ability.

[0108] Reasons for designing each item of the formula:

[0109] Ad embed vector :Ad embedding vector Used to represent the characteristics of advertisements, through the comprehensive vector The dot product measures the matching degree between the advertisement and the user’s characteristics and the current scene information, and can evaluate the fit between the advertisement and the user’s current interests and scenes.

[0110] Indicator function Used to capture specific attributes of an ad in a specific context, such as the type and content of the ad. Adjust the importance of different attributes to ensure that the model focuses on key attributes.

[0111] Interaction Functions It is used to calculate the interaction strength between the user feature vector and the current scene information. Adjust the importance of different feature interactions to capture the complex relationship between users and scenes.

[0112] Regularization term It is used to penalize the complexity of the user feature vector and the current scene information to prevent the model from overfitting. Control the strength of the regularization term to ensure that the model has good generalization ability.

[0113] The following is a brief introduction to how to obtain each parameter in the formula:

[0114] Ad embed vector It is obtained by training the ad recommendation model. The input of the model is ad features, and the output is ad embedding vector. Ad features can include ad category, content, title, picture, etc. Assume that the ad embedding vector obtained by training is .

[0115] User feature vector It is obtained by collecting historical behavior data such as user browsing history, search history, click history, etc., combined with user real-time behavior data, and processed by feature engineering. Assume that the user feature vector obtained by feature engineering is .

[0116] Current scene information The user's geographic location, time, weather conditions and other information are obtained in real time through the API interface. Assume that the current scene information obtained through the API interface is , indicates that the user is located in the city center, the time is afternoon, the weather is clear, and the user is using a mobile device.

[0117] Weight Matrix and and the bias vector and Obtained by training a neural network model, the model input is the user feature vector and current scene information, and the output is a comprehensive vector. Assume that the weight matrix and bias vector obtained through training are:

[0118]

[0119] Indicator function and weight factor Define the indicator function and weight factor according to business needs, obtained through expert knowledge or training model. Assume that the indicator function and weight factor are defined as:

[0120]

[0121]

[0122] Interaction Functions and weight factor By defining the interaction function and weight factor, combined with user characteristics and current scene information, the model is obtained through training. Assume that the interaction function and weight factor are defined as:

[0123]

[0124]

[0125] Regularization term and the regularization coefficient By defining the regularization term and regularization coefficient, the optimal value is determined through cross-validation and other techniques to balance the complexity and generalization ability of the model. Assume that the regularization coefficient determined by cross-validation is .

[0126] In the embodiment of the present application, it is assumed that we implement the above formula on an online advertising platform, and the specific parameters are as follows:

[0127] User feature vector

[0128] Current scene information

[0129] Ad embed vector

[0130] Weight Matrix

[0131] Weight Matrix

[0132] Bias Vector

[0133] Bias Vector

[0134] Indicator function and weight factor :

[0135]

[0136]

[0137] Interaction Functions and weight factor :

[0138]

[0139]

[0140] Regularization coefficient

[0141] Calculation process

[0142] Calculate the comprehensive vector :

[0143] Concatenate the user feature vector and the current scene information into a vector:

[0144] First layer calculation:

[0145]

[0146] Activation Function

[0147] Second layer calculation:

[0148]

[0149] Activation Function

[0150] Comprehensive vector

[0151] Calculate the dot product of the ad embedding vector and the synthesis vector:

[0152]

[0153] Calculate the indicator function part:

[0154]

[0155] Calculate the interactive function part:

[0156]

[0157] Calculate the regularization term:

[0158]

[0159] Calculate the final evaluation result:

[0160]

[0161] According to the calculation results, the evaluation result of the candidate ad in the current scenario is 1.3802. Assuming the threshold is set to 1, since the evaluation result is higher than the set threshold, it indicates that the ad has a higher expected performance in the current scenario, so it can be recommended to users first. In this way, the system can more accurately select target ads for push, increase the click-through rate and conversion rate of ads, and enhance user experience and commercial value.

[0162] Figure 2 A schematic diagram of the structure of an advertisement recommendation system based on a fusion neural network is provided for an embodiment of the present application. Figure 2 As shown, the device comprises:

[0163] Receiving module 21, used to receive real-time behavior data of users;

[0164] A construction module 22, configured to dynamically construct a user interest graph based on the real-time behavior data, wherein the user interest graph includes nodes of interest to the user and associations between nodes of interest;

[0165] An analysis module 23, configured to analyze the user interest graph in real time using a fusion neural network model, wherein the fusion neural network model is configured to predict the user's preference probability for different advertising categories by integrating the user's historical behavior data and current scene information;

[0166] A screening module 24, configured to screen out a set of highly relevant advertisements as candidate advertisements based on the preference probability;

[0167] The push module 25 is used to select a target advertisement from the candidate advertisements for push based on the current scenario of the user, so as to improve the click-through rate and conversion rate of the target advertisement; wherein the fusion neural network model is also used to optimize the advertisement recommendation strategy according to the feedback mechanism to ensure that the recommended target advertisement is highly matched with the user's interests.

[0168] Figure 2 The advertising recommendation system based on fusion neural network can be executed Figure 1 The implementation principle and technical effect of the advertising recommendation method based on a fusion neural network described in the embodiment shown will not be repeated. The specific way in which each module and unit performs operations in the advertising recommendation system based on a fusion neural network in the above embodiment has been described in detail in the embodiment of the method, and will not be elaborated here.

[0169] In one possible design, Figure 2 An advertisement recommendation system based on a fusion neural network of the embodiment shown can be implemented as a computing device, such as Figure 3 As shown, the computing device may include a storage component 31 and a processing component 32;

[0170] The storage component 31 stores one or more computer instructions, wherein the one or more computer instructions are called and executed by the processing component 32 .

[0171] The processing component 32 is used to: receive real-time behavior data of users; dynamically construct a user interest map based on the real-time behavior data, wherein the user interest map includes nodes of interest to the user and associations between nodes of interest; analyze the user interest map in real time using a fusion neural network model; screen out a set of highly relevant advertisements as candidate advertisements based on the preference probability; and select a target advertisement from the candidate advertisements for push based on the current scenario of the user, so as to improve the click-through rate and conversion rate of the target advertisement.

[0172] The processing component 32 may include one or more processors to execute computer instructions to complete all or part of the steps in the above method. Of course, the processing component may also be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors or other electronic components to perform the above method.

[0173] The storage component 31 is configured to store various types of data to support operations at the terminal. The storage component can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.

[0174] Of course, the computing device may also include other components, such as input / output interfaces, display components, communication components, etc.

[0175] The input / output interface provides an interface between the processing component and the peripheral interface module, which may be an output device, an input device, etc.

[0176] The communication component is configured to facilitate, among other things, wired or wireless communications between the computing device and other devices.

[0177] Among them, the computing device can be a physical device or an elastic computing host provided by a cloud computing platform, etc. In this case, the computing device can refer to a cloud server, and the above-mentioned processing components, storage components, etc. can be basic server resources rented or purchased from the cloud computing platform.

[0178] The present application also provides a computer storage medium storing a computer program, wherein the computer program can achieve the above-mentioned Figure 1 The illustrated embodiment is an advertisement recommendation method based on a fusion neural network.

[0179] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0180] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0181] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0182] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. An advertisement recommendation method based on fusion neural network, characterized in that: include: Receive real-time behavior data of users; Based on the real-time behavior data, dynamically construct a user interest graph, wherein the user interest graph includes nodes of interest to the user and associations between nodes of interest; Analyzing the user interest graph in real time using a fusion neural network model, wherein the fusion neural network model is used to predict the user's preference probability for different advertising categories by integrating the user's historical behavior data and current scene information; According to the preference probability, a set of highly relevant advertisements is screened out as candidate advertisements; In combination with the current scenario of the user, a target advertisement is selected from the candidate advertisements for push, so as to improve the click-through rate and conversion rate of the target advertisement; wherein the fusion neural network model is also used to optimize the advertisement recommendation strategy according to the feedback mechanism, so as to ensure that the recommended target advertisement is highly matched with the user's interests; The process of predicting the user's preference probability for a specific advertisement category by the fusion neural network model includes: The probability of the user's preference for a specific advertisement category is calculated using the following formula: ； in, Indicates that the user is at time For specific ad categories The probability of preference; It is the user's real-time behavior data and historical behavior data in time. The eigenvector of For specific ad categories The weight matrix of For specific ad categories The bias term of Is the user at time Current scene information, including geographic location, time, weather conditions, and device status; It is the weight matrix used to fuse the current scene information; ReLU function , used to fuse the user's feature vector and current scene information ; is the regularization coefficient, used to control the weight matrix The size of , to prevent overfitting; is the weight matrix The L2 norm squared; is the set of all advertising categories; For specific ad categories The bias of is a weight matrix for a specific ad category; in, Used to calculate the user's feature vector For specific ad categories Contribution Used to calculate the current scene information For specific ad categories Contribution Baseline used to adjust overall ratings; Used to prevent overfitting; as well as In Represents the transpose of the weight matrix.

2. The method according to claim 1, characterized in that The dynamically constructing a user interest graph based on the real-time behavior data includes: Extracting user behavior characteristics from the real-time behavior data, wherein the behavior characteristics include the user's browsing habits, purchasing tendencies, and social interactions; The behavior features are mapped to nodes in the user interest graph, and connecting edges are established between the nodes based on similarity or co-occurrence frequency to construct a user interest graph.

3. The method according to claim 1, characterized in that The real-time analysis of the user interest graph using a fusion neural network model includes: Inputting the real-time behavior data and the acquired historical behavior data of the user into a pre-trained fusion neural network model, so as to capture the nonlinear relationship in the user behavior and the interdependence between different behaviors by using a multi-level hidden layer in the fusion neural network model; Based on the nonlinear relationship and the interdependent relationship, and in combination with current scenario information, the user's preference probability for different types of advertisements is determined, wherein the current scenario information is used to increase the model's ability to predict user behavior in specific scenarios.

4. The method according to claim 3, characterized in that The step of selecting a set of highly relevant advertisements as candidate advertisements based on the preference probability includes: Sort the preference probabilities of different types of advertisements and select the advertisement categories whose preference probabilities are higher than a preset threshold; Building an advertisement library, wherein the advertisement library contains all optional advertisements and information corresponding to the optional advertisements, wherein the information corresponding to the optional advertisements includes advertisement category labels, advertisement content descriptions, and target audience group definitions; Based on a pre-built advertisement library index, the filtered advertisement categories are matched with optional advertisements in the advertisement library to find an advertisement set that matches the user preference, wherein the advertisement library index is used to match the optional advertisements in the advertisement library with the user preference according to advertisement category labels, content descriptions, and target audience group definitions; The matched advertisement set is sorted according to the comprehensive score of advertisement relevance and user preference, and advertisements with scores higher than a preset score are selected as candidate advertisements.

5. The method according to claim 4, characterized in that The selecting a target advertisement from the candidate advertisements for push in combination with the current scene of the user, so as to improve the click rate and conversion rate of the target advertisement, includes: Acquire scene information corresponding to the current scene where the user is located through the environment perception module, including geographic location, time, weather conditions, and device status; Analyzing the scene information, extracting key scene features, and combining the key scene features with the real-time behavior data to form comprehensive context data; Inputting the comprehensive context data into the fusion neural network model, evaluating the expected performance of the candidate advertisement in the current scenario, and obtaining an evaluation result; According to the evaluation result, select the target advertisement that best matches the current scenario and has the highest potential click rate and conversion rate from the candidate advertisements; Formulate a corresponding advertisement push strategy according to the target advertisement, wherein the advertisement push strategy includes a push time window, a push form, and a push channel; The advertisement push strategy is executed to improve the click-through rate and conversion rate of the target advertisement.

6. The method according to claim 1, characterized in that The fusion neural network model evaluates the candidate advertisement in the current scenario and the expected performance, and the process of obtaining the evaluation result includes: The candidate advertisement's expected performance in the current scenario is calculated using the following formula to obtain an evaluation result: ； in, Indicates at time and comprehensive contextual data Candidate ads under the assessment results; Is a candidate ad The embedding vector of ; is a nonlinear function used to integrate the user feature vector and current situation information , generating a comprehensive vector; is a series of indicator functions, each indicator function corresponds to a candidate ad In comprehensive contextual data An attribute in is each indicator function The weight factor is used to adjust the importance of attributes corresponding to different comprehensive context data; Is a set of functions used to calculate user feature vectors Scene information related to current situation information The strength of the interaction between Each function The weight factor is used to adjust the importance of the interaction between different features; It is a regularization term used to penalize the complexity of the user feature vector and the current context information to prevent overfitting; is the regularization term The coefficient of in, , Represents the user feature vector and current situation information Connect into a vector; and is the weight matrix; and is the bias vector; is the activation function.

7. An advertisement recommendation system based on fusion neural network, characterized in that: include: A receiving module, used to receive real-time behavior data of users; A construction module, used to dynamically construct a user interest graph based on the real-time behavior data, wherein the user interest graph includes nodes of interest to the user and associations between nodes of interest; An analysis module, used to analyze the user interest graph in real time using a fusion neural network model, wherein the fusion neural network model is used to predict the user's preference probability for a specific advertising category by integrating the user's historical behavior data and current scene information; A screening module, used to screen out a set of highly relevant advertisements as candidate advertisements based on the preference probability; A push module, which is used to select a target advertisement from the candidate advertisements for push in combination with the current scenario of the user, so as to improve the click-through rate and conversion rate of the target advertisement; wherein the fusion neural network model is also used to optimize the advertisement recommendation strategy according to the feedback mechanism, so as to ensure that the recommended target advertisement is highly matched with the user's interests; The process of predicting the user's preference probability for a specific advertisement category by the fusion neural network model includes: The probability of the user's preference for a specific advertisement category is calculated using the following formula: ； in, Indicates that the user is at time For specific ad categories The probability of preference; It is the user's real-time behavior data and historical behavior data in time. The eigenvector of For specific ad categories The weight matrix of For specific ad categories The bias term of Is the user at time Current scene information, including geographic location, time, weather conditions, and device status; It is the weight matrix used to fuse the current scene information; ReLU function , used to fuse the user's feature vector and current scene information ; is the regularization coefficient, used to control the weight matrix The size of , to prevent overfitting; is the weight matrix The L2 norm squared; is the set of all advertising categories; For specific ad categories The bias, is a weight matrix for a specific ad category; in, Used to calculate the user's feature vector For specific ad categories Contribution Used to calculate the current scene information For specific ad categories Contribution Baseline used to adjust overall ratings; Used to prevent overfitting; as well as In Represents the transpose of the weight matrix.

8. A computing device, characterized in that It comprises a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement an advertisement recommendation method based on a fusion neural network as described in any one of claims 1 to 6.

9. A computer storage medium, characterized in that A computer program is stored, and when the computer program is executed by a computer, an advertisement recommendation method based on a fusion neural network as described in any one of claims 1 to 6 is implemented.

Images