Advertisement information pushing method and system based on multi-source information

Abstract

The invention discloses an advertisement information pushing method and system based on multi-source information, and relates to the technical field of advertisement information pushing, and the method comprises the following components: S1, a data collection step, S2, an emotion and psychological state analysis step, S3, an advertisement matching step, S4, a VR / AR advertisement customization step, and S5, an advertisement pushing step. According to the method, social media speech, webpage browsing content and voice interaction record multi-source information of the user are acquired through multiple channels, and emotion and psychological state analysis is performed by using the deep neural network model, so that the current emotion state and potential psychological needs of the user can be accurately judged; the personalized advertisement information pushing mode remarkably improves the pertinence and effectiveness of the advertisement, so that the advertisement content is closer to the interest and demand of the user, the advertisement experience of the user is improved, and meanwhile, the advertisement dissensitivity of the user is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of advertising information push, and in particular to an advertising information push method and system based on multi-source information. Background Art

[0002] In today's information society, advertising information push has become an important bridge connecting consumers with goods or services. With the rapid development of big data and artificial intelligence technologies, the way of advertising push has gradually changed from the traditional wide-ranging method to precise push.

[0003] Traditional advertising push methods have obvious shortcomings. On the one hand, traditional methods often rely on a single data source, such as users' browsing history or purchase history, which leads to an incomplete and in-depth judgment of users' emotions and psychological states. On the other hand, traditional methods lack efficient algorithms and model support when analyzing user emotions, making it difficult to accurately capture users' potential needs and emotional changes, resulting in inaccurate advertising push and easily triggering users' disgust and resistance.

[0004] In summary, the traditional advertising push method has many limitations and is difficult to meet the personalized needs of the modern advertising market. Therefore, it is particularly important to develop an advertising information push method and system based on multi-source information. Summary of the invention

[0005] The purpose of the present invention is to make up for the shortcomings of the prior art and to provide an advertising information push method and system based on multi-source information. It can collect multi-source information such as users' social media speeches, web browsing content, and voice interaction records through multiple channels, and use a deep neural network model to perform emotion and psychological state analysis. It can more accurately judge the user's current emotional state and potential psychological needs, thereby pushing advertisements that are more in line with the user's interests and needs, thereby improving the effect of advertising push and user satisfaction.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions: a method for pushing advertising information based on multi-source information, the specific steps of the method are:

[0007] S1. Data collection steps: Collect multi-source information through multiple channels, including user social media speeches, web browsing content, voice interaction records, basic information, browsing history and purchase records. For social media speeches, natural language processing technology is used to extract text features through part-of-speech tagging and syntactic analysis operations. For web browsing content, web crawler technology is used to obtain page information, which is then processed by text parsing and feature extraction algorithms. Voice interaction records are converted into text with the help of speech recognition technology and then feature extraction is performed. The collected data is normalized and pre-processed to unify different types of data into the same value range;

[0008] S2. Emotion and psychological state analysis steps: Construct an emotion and psychological state analysis model based on a deep neural network structure, which includes an input layer, multiple hidden layers and an output layer. The input layer receives the preprocessed multi-source information feature vector, and the hidden layer uses an activation function where a 1 and a 2 It is a trainable parameter, which is dynamically adjusted during the training process through the back propagation algorithm and the stochastic gradient descent method. When training the model, a large amount of multi-source information data with labeled emotions and psychological states is used as the training set, and the cross entropy loss function is used. is the optimization objective, where y i is the true label, To predict the probability of the model, an attention mechanism is used between hidden layers to calculate the weight w of each source information j , the formula is where h j is the second source information feature, q is the query vector, and the score function is calculated based on the dot product or other similarity metrics. The model is used to analyze and judge the user's current emotional state and potential psychological needs;

[0009] S3, Advertisement matching step: Establish an advertisement database, classify advertisements by product type, applicable scenario, and emotional appeal, and use a matching algorithm to filter advertisements from the advertisement database based on the results of emotion and psychological state analysis. The matching algorithm calculates the similarity between the user feature vector and the advertisement feature vector based on cosine similarity. in is the user feature vector, is the ad feature vector, and a similarity threshold θ is set. Ads above the threshold are taken as matching results. Meanwhile, the preference degree of users’ historical behavior for different types of ads is considered, and the preference weight p is set. l , comprehensive calculation of matching score M = S × p l , filter ads by score from high to low;

[0010] S4. VR / AR advertising customization steps: Combine multi-source user information, use 3D modeling technology and AR development tools to customize VR / AR advertising. For car enthusiasts, determine the scene according to their geographic location, use the 3D model library to build a car model, combine the ray tracing algorithm to simulate the real lighting effect, and render a realistic car display scene. In AR advertising, use the feature point matching algorithm to merge the virtual car with the real scene, use the augmented reality registration algorithm to ensure the stable display of virtual objects in the real scene, and adjust the advertising content according to user interests;

[0011] S5. Advertisement push step: Select the push channel based on the user device information. When pushing, record the user's feedback on the advertisement, including whether it is clicked, browsing time, and interactive behavior. Use the feedback information to update the model and adjust the advertisement matching strategy to form a closed-loop optimization system.

[0012] Furthermore, in the data collection step, semantic role labeling technology is also used for feature extraction of social media speeches. Semantic role labeling aims to identify the semantic role of each predicate in a sentence. By labeling social media speeches with semantic roles, it is possible to have a deeper understanding of the meaning of the text and extract richer semantic features. These semantic role information are combined with part of speech and syntactic features to form a more comprehensive feature vector of social media speeches, providing a more accurate data basis for subsequent emotion and psychological state analysis. When normalizing and preprocessing data from different sources, different normalization methods are used according to the characteristics of different types of data. For numerical data, the maximum-minimum normalization method is used to map the data to the [0, 1] interval. The formula is: Where x is the original data, x min and x m ax are the minimum and maximum values ​​of the data type respectively. For text data, after the text is converted into a vector through the word vector model, the 12 normalization method is used to make the vector modulus length 1. The formula is: in is the original text vector to ensure that different types of data are analyzed at the same scale.

[0013] Furthermore, in the emotion and psychological state analysis step, the activation function Parameter a 1 and a 2 The method for determining is as follows: at the beginning of model training, a is randomly initialized 1 and a 2 The value of is between (0, 1). During the training process, the back propagation algorithm is used to calculate the loss function of a 1 and a 2 The gradient of a is updated by stochastic gradient descent 1 and a 2 The specific update formula is Where η is the learning rate, t is the training round, and by continuously adjusting a 1 and a 2 The value of , which gradually reduces the loss function of the model on the training set, thereby optimizing the model performance. At the same time, in order to avoid overfitting, a regularization method is used in the training process, 12 regularization, for a 1 and a 2 Constraint, the formula is L regnularized =L+λ(|a1 | 2 +|a 2 | 2 ), where λ is the regularization coefficient, and its optimal value is determined through experiments to balance the fitting ability and generalization ability of the model.

[0014] Furthermore, in the emotion and psychological state analysis step, the query vector q in the attention mechanism is generated as follows. The query vector q is obtained by linearly transforming the multi-source information feature vector of the input layer. Specifically, the multi-source information feature vector of the input layer is assumed to be Concatenate these eigenvectors into a large eigenmatrix Then through a trainable weight matrix W q Perform matrix multiplication with the feature matrix X to obtain the query vector q = W q X, where W q The dimension of is determined according to the input feature dimension and the expected query vector dimension. During the training process, the weight matrix W q The query vector q generated by the back-propagation algorithm and the stochastic gradient descent method is updated to optimize the effect of the attention mechanism. The query vector q generated in this way can comprehensively consider the characteristics of multi-source information, thereby more accurately calculating the weight w of each source information. j , improve the accuracy of emotion and psychological state analysis.

[0015] Furthermore, in the advertisement matching step, the method for constructing the advertisement feature vector is as follows: for each advertisement, first, different initial feature values ​​are assigned according to its classification information. For an advertisement for a relaxation product for anxious people, a higher initial value is assigned in the emotional appeal dimension. Then, the text feature is extracted from the advertisement copy, and the bag-of-words model or the word vector model is used to convert the copy into a vector form. For the image information in the advertisement, a convolutional neural network is used to extract features to obtain an image feature vector. The text feature vector and the image feature vector are concatenated, and then a fully connected layer is used for dimensionality reduction and feature fusion to obtain the final advertisement feature vector. In the process of constructing the advertising feature vector, the historical advertising effect data is also taken into consideration and incorporated into the advertising feature vector as additional features to improve the accuracy of advertising matching.

[0016] Furthermore, in the advertisement matching step, the preference weights p of the user's historical behaviors for different types of advertisements are l The method for determining is as follows: statistics are collected on the number of clicks, browsing time, and purchase behavior of users on different types of ads in the past period of time. Different weights are assigned to the number of clicks according to the time of the clicks. Recent clicks have higher weights, while long-term clicks have lower weights. The calculation formula is: where ci is the number of clicks for the i-th time, t i is the number of days from the current time, β is the attenuation coefficient, and the browsing time is also weighted according to the time distance. The formula is: where v i is the duration of the i-th browsing, s i is the number of days from the current time. A higher weight is given to purchase behavior, which is calculated based on the purchase amount and number of times. The formula is: where a j is the amount n of the jth purchase j is the number of purchases, u j is the number of days from the current time, and finally p click、 p view and p purchase Perform weighted summation to obtain the preference weight p l , the formula is p l =w 1 p click +w 2 p view +w 3 p purchase , where w 1 、w 2 and w 3 is the weight coefficient, and its optimal value is determined through experiments.

[0017] Furthermore, in the VR / AR advertising customization step, the ray tracing algorithm adopts a ray propagation model when simulating real lighting effects. Traditional ray tracing algorithms usually only consider direct lighting when calculating the interaction between light and the surface of an object. The model of the present invention adds the calculation of indirect lighting. Specifically, when light intersects the surface of an object, in addition to calculating direct reflection and refracted light, the Monte Carlo method is used to randomly sample light in the surrounding environment to calculate the indirect lighting contribution of these light rays to the surface of the object. For each sampled light ray, the propagation and reflection of the light in the scene are recursively calculated by calculating its intersection with the surrounding objects. When calculating the indirect lighting contribution, the material properties of the object surface are taken into account to more realistically simulate the reflection and absorption characteristics of different materials to light. Through this improved ray propagation model, more realistic lighting effects can be rendered in VR / AR advertisements, enhancing the user's immersive experience.

[0018] Furthermore, in the VR / AR advertisement customization step, the augmented reality registration algorithm adopts a method of fusion of feature points and depth information, uses a feature point detection algorithm to extract feature points in the real scene image, and establishes a feature point descriptor, obtains the depth information of the real scene through a depth sensor, and when the virtual object is fused with the real scene, the position and posture of the virtual object in the real scene are preliminarily determined according to the feature point matching result, and the position and posture of the virtual object are further optimized in combination with the depth information. Specifically, by comparing the depth information d of the virtual object v Depth information d corresponding to the position in the real scene r , adjust the position of the virtual object to make it more matched with the real scene in the depth direction. In terms of posture adjustment, the relative angle θ between the virtual object and the object in the real scene is calculated using the depth information, and the posture of the virtual object is fine-tuned to ensure the stable display and realism of the virtual object in the real scene. Through this augmented reality registration algorithm based on the fusion of feature points and depth information, the fusion accuracy of virtual objects and real scenes can be improved, providing users with a better AR advertising experience.

[0019] On the other hand, an advertising information push system based on multi-source information is characterized in that the system includes a data acquisition module, an emotion and psychological state analysis module, an advertising matching module and an advertising push module:

[0020] The data collection module collects multi-source information through multiple channels, including user social media speeches, web browsing content, voice interaction records, basic information, browsing history and purchase records. For social media speeches, natural language processing technology is used to extract text features through part-of-speech tagging and syntactic analysis operations. For web browsing content, web crawler technology is used to obtain page information, which is then processed by text parsing and feature extraction algorithms. Voice interaction records are converted into text with the help of speech recognition technology and then feature extraction is performed. The collected data is normalized and pre-processed to unify different types of data into the same value range.

[0021] The emotion and psychological state analysis module: constructs an emotion and psychological state analysis model, which is based on a deep neural network structure and includes an input layer, multiple hidden layers and an output layer. The input layer receives the preprocessed multi-source information feature vector, and the hidden layer uses an activation function where a 1 and a 2 It is a trainable parameter, which is dynamically adjusted during the training process through the back propagation algorithm and the stochastic gradient descent method. When training the model, a large amount of multi-source information data with labeled emotions and psychological states is used as the training set, and the cross entropy loss function is used. is the optimization objective, where y iis the true label, To predict the probability of the model, an attention mechanism is used between hidden layers to calculate the weight w of each source information j , the formula is where h j is the second source information feature, q is the query vector, and the score function is calculated based on the dot product or other similarity metrics. The model is used to analyze and judge the user's current emotional state and potential psychological needs;

[0022] The advertisement matching module: establishes an advertisement database, classifies advertisements by product type, applicable scenario, and emotional appeal, and uses a matching algorithm to filter advertisements from the advertisement database based on the results of emotion and psychological state analysis. The matching algorithm calculates the similarity between the user feature vector and the advertisement feature vector based on cosine similarity. in is the user feature vector, is the ad feature vector, and a similarity threshold θ is set. Ads above the threshold are taken as matching results. Meanwhile, the preference degree of users’ historical behavior for different types of ads is considered, and the preference weight p is set. l , comprehensive calculation of matching score M = S × p l , filter ads by score from high to low;

[0023] The advertisement matching module: combines multi-source information of users, uses 3D modeling technology and AR development tools to customize VR / AR advertisements. For car enthusiasts, the scene is determined according to their geographic location, and a car model is constructed using a 3D model library. The ray tracing algorithm is used to simulate the real lighting effect and render a realistic car display scene. In AR advertisements, the virtual car is integrated with the real scene through a feature point matching algorithm, and an augmented reality registration algorithm is used to ensure the stable display of virtual objects in the real scene, and the advertisement content is adjusted according to the user's interests.

[0024] The advertising push module selects a push channel based on user device information, and during push, records user feedback on the advertisement, including whether the advertisement is clicked, browsing time, and interactive behavior, and uses the feedback information for model update and advertisement matching strategy adjustment to form a closed-loop optimization system.

[0025] Compared with the prior art, the method and system for pushing advertising information based on multi-source information has the following beneficial effects:

[0026] 1. The present invention collects multi-source information such as users' social media speeches, web browsing content, and voice interaction records through multiple channels, and uses a deep neural network model to analyze emotions and psychological states. It can accurately judge the user's current emotional state and potential psychological needs. This personalized advertising information push method significantly improves the pertinence and effectiveness of advertising, making the advertising content more in line with the user's interests and needs, thereby enhancing the user's advertising experience and reducing the user's aversion to advertising.

[0027] 2. The present invention combines VR / AR technology to customize advertisements. Through 3D modeling and ray tracing algorithm technology, it can present realistic advertising scenes and interactive experiences to users. This innovative form of advertising not only enhances the attractiveness and fun of advertisements, but also improves user participation and interactivity. At the same time, it also updates the model and adjusts the advertisement matching strategy based on user feedback information, forming a closed-loop optimization system, further improving the effect of advertisement push and user satisfaction. This method and system for pushing advertisement information based on multi-source information provides advertisers with a more efficient and accurate way of advertising delivery, and also brings users a better advertising experience.

[0028] Other advantages, objectives and features of the present invention will be set forth in part in the following description and, in part, will be apparent to those skilled in the art based on an examination of the following or may be taught from the practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0030] Figure 1 It is a flow chart of an advertising information push method based on multi-source information;

[0031] Figure 2 The present invention is a process operation diagram of an advertising information push system based on multi-source information. DETAILED DESCRIPTION

[0032] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the specific implementation mode, structure, characteristics and effects of the present invention are described in detail below in combination with the accompanying drawings and preferred embodiments.

[0033] Embodiment 1

[0034] This embodiment describes the push of advertisements for car enthusiasts.

[0035] For car enthusiasts, their speeches are collected through social media platforms, such as evaluations of different car models and sharing of modification experiences on car forums. Natural language processing technology is used for part-of-speech tagging and syntactic analysis, and semantic role labeling technology is used to deeply understand the meaning of the text and extract rich semantic features. For example, the semantic roles related to the predicates "recommend" and "like" in the sentence are identified to judge their preference for cars.

[0036] Web crawlers are used to obtain the car-related web page content that they browse, such as car review websites and official car brand websites. Through text parsing and feature extraction algorithm processing, the page information is converted into data that can be analyzed.

[0037] Collect voice interaction records, such as the voice content of car maintenance knowledge asked by the smart voice assistant, and convert it into text with the help of speech recognition technology for feature extraction.

[0038] Integrate their basic information (such as age, gender, occupation), browsing history (browsed car pictures, videos, articles) and purchase records (whether they have purchased car-related products), and normalize and preprocess the collected data. The maximum-minimum normalization method is used for numerical data. For example, the browsing frequency of car-related content in the browsing history is mapped to the [0, 1] interval. After the text data is converted into a vector through the word vector model, the L2 normalization method is used to make the vector modulus length 1.

[0039] Construct an emotion and psychological state analysis model based on a deep neural network. The input layer receives the preprocessed multi-source information feature vector, and the hidden layer uses an activation function Initialize the model randomly at the beginning of training 1 and a 2 The value of is between (0, 1) and the back propagation algorithm is used to calculate the loss function during training. 1 and a 2 The gradient of is updated by stochastic gradient descent method, and the update formula is (where η is the learning rate and t is the training round), and the L2 regularization method is used to 1 and a 2 Constraint, the formula is L r egmularized=L+λ(|a 1 | 2 +|a 2 | 2 )(λ is the regularization coefficient), the model training uses a large amount of multi-source information data labeled with emotions and psychological states as the training set, and the cross entropy loss function To optimize the target, an attention mechanism is used between hidden layers. The query vector q is obtained by linearly transforming the multi-source information feature vector of the input layer. Let the input layer feature vector be x 1 ,x 2 ,…,x m , concatenated into a feature matrix X = [x 1 ,x 2 ,…,x m ], then q = W q X(W q is a trainable weight matrix), calculate the weight of each source information (h j is the jth source information feature, and the score function is calculated based on the dot product or other similarity metrics to analyze and judge the user's current emotional state (such as the degree of love for cars, expectations for specific models) and potential psychological needs (such as the pursuit of high-performance cars, interest in car modification).

[0040] Establish an advertising database, classify automobile-related advertisements by product type (such as vehicle sales, auto parts, and auto maintenance services), applicable scenarios (such as urban driving and off-road scenarios), and emotional appeal dimensions (such as emphasizing speed and passion, comfort and luxury). For each advertisement, assign different initial feature values ​​according to its classification information. For example, a higher initial value is assigned to the emotional appeal dimension for automobile advertisements that pursue extreme passion. The advertising copy is converted into a vector form using a bag-of-words model or a word vector model. The image information in the advertisement is extracted using a convolutional neural network. The text and image feature vectors are concatenated and then subjected to dimensionality reduction and feature fusion by a fully connected layer to obtain the final advertising feature vector a. At the same time, the historical advertising delivery effect data is incorporated into the advertising feature vector. According to the results of the emotion and psychological state analysis, a matching algorithm is used to calculate the similarity between the user feature vector u and the advertising feature vector a based on cosine similarity. A similarity threshold θ is set, and advertisements above the threshold are taken as preliminary matching results. At the same time, the number of clicks on different types of automobile advertisements by users in the past period of time is counted (the number of clicks is weighted according to the time distance, is the number of clicks for the i-th time, t i is the number of days from the current time, β is the decay coefficient), browsing time ( v i is the duration of the i-th browsing, s i is the number of days from the current time), purchase behavior ( a j is the jth purchase amount, n j is the number of purchases u j is the number of days from the current time), and the weighted sum of these data is obtained to obtain the preference weight p l =w 1 pclick +w 2 p view +w 3 p purchase (w 1 、w 2 、w 3 is the weight coefficient), and the comprehensive calculation matching score M = S × p l , filter ads by score from high to low.

[0041] Combine multi-source user information, use 3D modeling technology and AR development tools to customize VR / AR ads, and determine scenes based on the geographic location of car enthusiasts, such as building scenes near car modification shops or around racetracks where they often visit. Use 3D model libraries to build car models, and combine ray tracing algorithms to simulate real lighting effects. The ray tracing algorithm uses a light propagation model and adds indirect lighting calculations. In addition to calculating direct reflection and refraction light when light intersects the surface of an object, the Monte Carlo method randomly samples ambient light to calculate indirect lighting contributions, and takes into account the material properties of the object surface.

[0042] In AR advertising, the virtual car is integrated with the real scene through the feature point matching algorithm, the feature point and depth information fusion method is adopted by the augmented reality registration algorithm, the feature point detection algorithm is used to extract the feature points in the real scene image and establish descriptors, the depth information is obtained through the depth sensor, the position and posture of the virtual object are preliminarily determined according to the feature point matching results, and further optimized in combination with the depth information, and the advertising content is adjusted according to the user's interests. For example, if the user is interested in a specific car color or modification style, the relevant content will be highlighted in the advertisement.

[0043] Select push channels based on user device information (such as mobile phone model, operating system), such as push to the user's car-related APP or mobile phone text messages. Record user feedback on ads during push, including whether clicked, browsing time, interactive behavior (such as sharing, liking), and use feedback information for model update and ad matching strategy adjustment. For example, based on the user's click behavior on a specific ad, adjust the ad push strategy for the user and increase the push frequency of similar ads; optimize the emotion and psychological state analysis model and ad matching algorithm parameters based on the user's browsing time and interactive behavior to form a closed-loop optimization system.

[0044] Embodiment 2

[0045] This embodiment describes the push of advertisements for travel enthusiasts.

[0046] Collect travel experiences and destination recommendations shared by travel enthusiasts from social media platforms, and use natural language processing technology to extract features, such as descriptions of attractions and travel experiences.

[0047] Through web crawlers, we obtain the web content of travel guide websites and official websites of tourist attractions that they browse, and parse and extract relevant information.

[0048] Collect voice interaction records, such as asking a voice assistant about travel route planning, and convert them into text to extract features.

[0049] Their basic information, browsing history (browsed travel photos and videos) and purchase records (such as purchased travel equipment and booked hotels) are sorted out, and the data is normalized and preprocessed. Corresponding normalization methods are used for numerical data and text data respectively.

[0050] A deep neural network model is constructed, and the input layer receives the preprocessed multi-source information feature vector. The hidden layer activation function and parameter adjustment and training process are similar to those in Example 1. A large amount of labeled data is used for training, and the cross entropy loss function is optimized. The attention mechanism is used to calculate the weight of each source information. The model analyzes the user's emotional state of travel (such as the degree of love for different destinations, satisfaction in travel) and potential psychological needs (such as the desire for adventure travel and the pursuit of cultural experience).

[0051] Establish a tourism-related advertising database, with classification methods such as tourism destination type (seaside, mountainous area), tourism product type (group tours, free travel), and emotional appeal (relaxation and leisure, exploration and adventure).

[0052] Construct an advertising feature vector, consider the text, image information and historical delivery effect data, calculate the similarity between the user feature vector and the advertising feature vector, combine the user's historical behavior preference weight, and comprehensively calculate the matching score to filter advertisements. For example, if the user has clicked on seaside travel advertisements many times in the past and has recently browsed seaside travel guides, the system will give priority to pushing advertisements related to seaside travel destinations.

[0053] Customize VR / AR travel ads based on user information. For example, build virtual scenes based on tourist destinations that users frequently visit, use 3D modeling technology to build models of local famous attractions, and use ray tracing algorithms to simulate real lighting. In AR ads, merge virtual attractions with the user's real scene. Augmented reality registration algorithms ensure stable display of virtual objects. Adjust advertising content based on user interests. For example, if a user likes historical and cultural attractions, the historical and cultural background of the attraction will be introduced in detail in the advertisement.

[0054] Advertisements are pushed based on user devices, and feedback information is recorded for optimization of models and strategies. For example, if a user clicks on an island tourism ad and browses it for a long time, the system will add similar island tourism ads or related tourism product ads in subsequent push notifications. At the same time, the sentiment analysis model and ad matching algorithm are optimized based on feedback information to improve push accuracy.

[0055] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Although the present invention has been disclosed as a preferred embodiment as above, it is not used to limit the present invention. Any technical personnel in this field can make some changes or modifications to the technical contents disclosed above without departing from the scope of the technical solution of the present invention. However, any brief modifications, different changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. A method for pushing advertising information based on multi-source information, characterized in that: The specific steps of this method are: S1. Data collection steps: Collect multi-source information through multiple channels, including user social media speeches, web browsing content, voice interaction records, basic information, browsing history and purchase records. For social media speeches, natural language processing technology is used to extract text features through part-of-speech tagging and syntactic analysis operations. For web browsing content, web crawler technology is used to obtain page information, which is then processed by text parsing and feature extraction algorithms. Voice interaction records are converted into text with the help of speech recognition technology and then feature extraction is performed. The collected data is normalized and pre-processed to unify different types of data into the same value range; S2. Emotion and psychological state analysis steps: Construct an emotion and psychological state analysis model based on a deep neural network structure, which includes an input layer, multiple hidden layers and an output layer. The input layer receives the preprocessed multi-source information feature vector, and the hidden layer uses an activation function Among them, a1 and a2 are trainable parameters, which are dynamically adjusted during the training process through the back propagation algorithm and the stochastic gradient descent method. When training the model, a large amount of multi-source information data with labeled emotions and psychological states is used as the training set, and the cross entropy loss function is used. is the optimization objective, where y i is the true label, To predict the probability of the model, an attention mechanism is used between hidden layers to calculate the weight w of each source information j , the formula is where h j is the second source information feature, q is the query vector, and the score function is calculated based on the dot product or other similarity metrics. The model is used to analyze and judge the user's current emotional state and potential psychological needs; S3, Advertisement matching step: Establish an advertisement database, classify advertisements by product type, applicable scenario, and emotional appeal, and use a matching algorithm to filter advertisements from the advertisement database based on the results of emotion and psychological state analysis. The matching algorithm calculates the similarity between the user feature vector and the advertisement feature vector based on cosine similarity. in is the user feature vector, is the ad feature vector, and a similarity threshold θ is set. Ads above the threshold are taken as matching results. Meanwhile, the preference degree of users’ historical behavior for different types of ads is considered, and the preference weight p is set. l , comprehensive calculation of matching score M = S × p l , filter ads by score from high to low; S4. VR / AR advertising customization steps: Combine multi-source user information, use 3D modeling technology and AR development tools to customize VR / AR advertising. For car enthusiasts, determine the scene according to their geographic location, use the 3D model library to build a car model, combine the ray tracing algorithm to simulate the real lighting effect, and render a realistic car display scene. In AR advertising, use the feature point matching algorithm to merge the virtual car with the real scene, use the augmented reality registration algorithm to ensure the stable display of virtual objects in the real scene, and adjust the advertising content according to user interests; S5. Advertisement push step: Select the push channel based on the user device information. When pushing, record the user's feedback on the advertisement, including whether it is clicked, browsing time, and interactive behavior. Use the feedback information to update the model and adjust the advertisement matching strategy to form a closed-loop optimization system.

2. The method for pushing advertising information based on multi-source information according to claim 1, characterized in that: In the data collection step, semantic role labeling technology is also used for feature extraction of social media speeches. Semantic role labeling aims to identify the semantic role of each predicate in a sentence. By labeling social media speeches with semantic roles, the meaning of the text can be more deeply understood and richer semantic features can be extracted. When normalizing and preprocessing data from different sources, different normalization methods are used according to the characteristics of different types of data. For numerical data, the maximum-minimum normalization method is used to map the data to the [0, 1] interval. The formula is: Where x is the original data, x min and x m ax are the minimum and maximum values ​​of the data type respectively. For text data, after the text is converted into a vector through the word vector model, the 12 normalization method is used to make the vector modulus length 1. The formula is: in is the original text vector.

3. The method for pushing advertising information based on multi-source information according to claim 1, characterized in that: In the emotion and psychological state analysis step, the activation function The method for determining the parameters a1 and a2 is as follows: at the beginning of model training, the values ​​of a1 and a2 are randomly initialized to a range of (0, 1). During the training process, the back propagation algorithm is used to calculate the gradient of the loss function with respect to a1 and a2, and the values ​​of a1 and a2 are updated by the stochastic gradient descent method. The specific update formula is: Where η is the learning rate, t is the training round, and by continuously adjusting the values ​​of a1 and a2, the loss function of the model on the training set is gradually reduced, thereby optimizing the model performance. At the same time, in order to avoid overfitting, a regularization method is used in the training process, 12 regularization, to constrain a1 and a2, the formula is L regnularized =L+λ(|a1| 2 +|a2| 2 ), where λ is the regularization coefficient, and its optimal value is determined through experiments to balance the fitting ability and generalization ability of the model.

4. The method for pushing advertising information based on multi-source information according to claim 1, characterized in that: In the emotion and psychological state analysis step, the query vector q in the attention mechanism is generated as follows. The query vector q is obtained by linearly transforming the multi-source information feature vector of the input layer. Specifically, the multi-source information feature vector of the input layer is Concatenate these eigenvectors into a large eigenmatrix Then through a trainable weight matrix W q Perform matrix multiplication with the feature matrix X to obtain the query vector q = W q X, where W q The dimension of is determined according to the input feature dimension and the expected query vector dimension. During the training process, the weight matrix W q The query vector q generated by the back propagation algorithm and the stochastic gradient descent method can comprehensively consider the characteristics of multi-source information, thereby more accurately calculating the weight w of each source information. j .

5. The method for pushing advertising information based on multi-source information according to claim 1, characterized in that: In the advertisement matching step, the method for constructing the advertisement feature vector is as follows: for each advertisement, first, different initial feature values ​​are assigned according to its classification information. For an advertisement for a relaxation product for anxious people, a higher initial value is assigned in the dimension of emotional appeal. Then, the text feature of the advertisement copy is extracted, and the copy is converted into a vector form using a bag-of-words model or a word vector model. For the image information in the advertisement, a convolutional neural network is used to extract features to obtain an image feature vector. The text feature vector and the image feature vector are concatenated, and then a fully connected layer is used for dimensionality reduction and feature fusion to obtain the final advertisement feature vector. In the process of constructing the advertising feature vector, the historical advertising effect data is also taken into consideration and incorporated into the advertising feature vector as additional features.

6. The method for pushing advertisement information based on multi-source information according to claim 1, characterized in that: In the ad matching step, the preference weights p of the user's historical behavior for different types of ads are l The method for determining is as follows: statistics are collected on the number of clicks, browsing time, and purchase behavior of users on different types of ads in the past period of time. Different weights are assigned to the number of clicks according to the time of the clicks. Recent clicks have higher weights, while long-term clicks have lower weights. The calculation formula is: where c i is the number of clicks for the i-th time, t i is the number of days from the current time, β is the attenuation coefficient, and the browsing time is also weighted according to the time distance. The formula is: where v i is the duration of the i-th browsing, s i is the number of days from the current time. A higher weight is given to purchase behavior, which is calculated based on the purchase amount and number of times. The formula is: where a j is the amount n of the jth purchase j is the number of purchases, u j is the number of days from the current time, and finally p click、 p view and p purchase Perform weighted summation to obtain the preference weight p l , the formula is p l =w1p click +w2p view +w3p purchase , where w1, w2 and w3 are weight coefficients.

7. The method for pushing advertisement information based on multi-source information according to claim 1, characterized in that: In the VR / AR advertisement customization step, the ray tracing algorithm adopts a ray propagation model when simulating the real lighting effect. The model of the present invention adds the calculation of indirect lighting. Specifically, when the light intersects with the surface of an object, in addition to calculating the direct reflection and refraction light, the Monte Carlo method is used to randomly sample the light in the surrounding environment, and the indirect lighting contribution of these light rays to the surface of the object is calculated. For each sampled light ray, the intersection with the surrounding objects is calculated, and the propagation and reflection of the light in the scene are recursively calculated. When calculating the indirect lighting contribution, the material properties of the object surface are taken into account.

8. The method for pushing advertisement information based on multi-source information according to claim 1, characterized in that: In the VR / AR advertisement customization step, the augmented reality registration algorithm adopts a method of fusion of feature points and depth information, uses a feature point detection algorithm to extract feature points in the real scene image, and establishes a feature point descriptor, and obtains the depth information of the real scene through a depth sensor. When the virtual object is fused with the real scene, the position and posture of the virtual object in the real scene are preliminarily determined based on the feature point matching results, and the position and posture of the virtual object are further optimized in combination with the depth information.

9. An advertising information push system based on multi-source information, characterized in that: The system includes a data collection module, an emotion and psychological state analysis module, an advertisement matching module, and an advertisement push module: The data collection module collects multi-source information through multiple channels, including user social media speeches, web browsing content, voice interaction records, basic information, browsing history and purchase records. For social media speeches, natural language processing technology is used to extract text features through part-of-speech tagging and syntactic analysis operations. For web browsing content, web crawler technology is used to obtain page information, which is then processed by text parsing and feature extraction algorithms. Voice interaction records are converted into text with the help of speech recognition technology and then feature extraction is performed. The collected data is normalized and pre-processed to unify different types of data into the same value range. The emotion and psychological state analysis module: constructs an emotion and psychological state analysis model, which is based on a deep neural network structure and includes an input layer, multiple hidden layers and an output layer. The input layer receives the preprocessed multi-source information feature vector, and the hidden layer uses an activation function Among them, a1 and a2 are trainable parameters, which are dynamically adjusted during the training process through the back propagation algorithm and the stochastic gradient descent method. When training the model, a large amount of multi-source information data with labeled emotions and psychological states is used as the training set, and the cross entropy loss function is used. is the optimization objective, where y i is the true label, To predict the probability of the model, an attention mechanism is used between hidden layers to calculate the weight w of each source information j , the formula is where h j is the second source information feature, q is the query vector, and the score function is calculated based on the dot product or other similarity metrics. The model is used to analyze and judge the user's current emotional state and potential psychological needs; The advertisement matching module: establishes an advertisement database, classifies advertisements by product type, applicable scenario, and emotional appeal, and uses a matching algorithm to filter advertisements from the advertisement database based on the results of emotion and psychological state analysis. The matching algorithm calculates the similarity between the user feature vector and the advertisement feature vector based on cosine similarity. in is the user feature vector, is the ad feature vector, and a similarity threshold θ is set. Ads above the threshold are taken as matching results. Meanwhile, the preference degree of users’ historical behavior for different types of ads is considered, and the preference weight p is set. l , comprehensive calculation of matching score M = S × p l , filter ads by score from high to low; The advertisement matching module: combines multi-source information of users, uses 3D modeling technology and AR development tools to customize VR / AR advertisements. For car enthusiasts, the scene is determined according to their geographic location, and a car model is constructed using a 3D model library. The ray tracing algorithm is used to simulate the real lighting effect and render a realistic car display scene. In AR advertisements, the virtual car is integrated with the real scene through a feature point matching algorithm, and an augmented reality registration algorithm is used to ensure the stable display of virtual objects in the real scene, and the advertisement content is adjusted according to the user's interests. The advertising push module selects a push channel based on user device information, and during push, records user feedback on the advertisement, including whether the advertisement is clicked, browsing time, and interactive behavior, and uses the feedback information for model update and advertisement matching strategy adjustment to form a closed-loop optimization system.

Images