Media delivery method and system based on water purification equipment data analysis

Abstract

The application discloses a media delivery method and system based on water purification equipment data analysis, relates to the technical field of media delivery, and divides an advertisement resource set into multiple class resource sets according to business information through a gradient boosting decision tree, links all the class resource sets according to business association to construct an advertisement resource library; information text of a target community is acquired based on equipment installation information, is classified into a corresponding class resource set by adopting a text classification algorithm, and a horizontal migration recommendation algorithm or random extraction is adopted to select a recommended advertisement for the target community based on whether or not there is an advertisement interaction record of the same type community; drinking water statistical data is acquired by counting and collecting water taking records in each adjustment period, and a personnel scale degree is calculated by fuzzy reasoning to determine a similar community; an advertisement is newly recommended for the target community by fusing and filtering a recommendation algorithm and considering the advertisement interaction records of the target community and the similar community, and effective combination of water purification equipment data and advertisement media is realized, and advertisement delivery precision and effect are improved.

Description

Technical Field

[0001] This invention relates to the field of media placement technology, specifically to a media placement method and system based on data analysis from water purification equipment. Background Technology

[0002] Traditional advertising methods lack precision, and a large amount of advertising resources are wasted on ineffective exposure, failing to truly reach the target customer group, resulting in poor advertising performance and an unsatisfactory return on investment.

[0003] With the rapid development of IoT technology, water purification equipment has become widely used in various companies and locations. The drinking water data and equipment installation information of water purification equipment can potentially reflect the company's staff size and business information, which is very beneficial for targeted advertising. Therefore, there is an urgent need for a media placement method and system based on water purification equipment data analysis to achieve precise advertising based on drinking water data and equipment installation information. Summary of the Invention

[0004] This invention addresses the shortcomings of existing technologies by proposing a media delivery method and system based on water purification equipment data analysis, in order to achieve intelligent advertising recommendations based on water purification equipment data.

[0005] The technical solution to achieve the purpose of this invention is as follows:

[0006] The media placement method based on water purification equipment data analysis includes the following specific steps:

[0007] Get advertising resource set And by using a gradient boosting decision tree, it is divided into categories based on business information. Each type of resource set is linked in a graph structure based on business relationships. Individual resource sets are used to build an advertising resource library. Total number of advertising categories;

[0008] Based on device installation information, the system retrieves text information from the target community via the internet and categorizes it into corresponding resource sets using a text classification algorithm. The system then determines whether to use a lateral migration recommendation algorithm or random sampling to select and recommend content to the target community based on the presence or absence of advertising interaction records from similar communities. An advertisement, This refers to the number of ads recommended in a single instance.

[0009] Collect and store water collection records and advertising interaction records of the target community. Collect drinking water statistics within each adjustment cycle. Analyze the drinking water statistics using fuzzy inference to calculate population size and identify similar communities. Then, use a fusion filtering recommendation algorithm to consider advertising interaction records between the target community and similar communities to re-recommend the target community. An advertisement.

[0010] Furthermore, advertising resource collection The advertising feature vectors in the data can be extracted through natural language processing technology by analyzing the business information of the advertisements, and then further input into the system. Decision trees generated by gradient boosting decision tree algorithm in rounds of pre-training Based on the Softmax function, the first Each advertising feature vector In decision tree China regarding The scores of each class are transformed into a probability distribution to obtain the scores of the classes belonging to the first class. class probability ,in, For the first Each advertising feature vector Class tags, For advertising feature vectors; the first Each advertising feature vector Predictive tags Select as The class with the highest probability among all classes. , , , and These represent the total number of ads, the total number of training rounds of the decision tree, and the total number of ad classes, respectively. After processing using the gradient boosting decision tree algorithm, the ad resource set is... Divided into Individual resource sets.

[0011] Furthermore, gradient boosting decision trees train a base learner in each iteration to fit the residual from the previous iteration, and then weighted and accumulated. The decision tree is obtained by iterative base learners. , The pre-training of a gradient boosting decision tree, which determines the total number of training rounds, includes the following specific steps:

[0012] Get by The advertising sample set consists of the feature vectors of each sample advertisement and the corresponding sample class labels. , The total number of sample advertisements;

[0013] Initialize constant predictor And satisfy the first Each sample advertisement feature vector After constant predictor The information obtained The scores of all classes are the same. Total number of advertising categories;

[0014] In the In the round of iteration, , For the total number of training rounds of the decision tree, for the th Each sample advertisement feature vector In the Decision tree generated in rounds of iteration Classification loss in Solve the problem about the decision tree separately. middle The partial derivatives of the scores of each class are taken, and their negative values ​​are used as the corresponding negative gradients.

[0015] Combination No. Each sample advertisement feature vector In the In each iteration, they belong to The negative gradient of the i-th class is used to construct the i-th class. Each sample advertisement feature vector In the Residual in round iteration Training the first Base learner in round iteration Defit The residuals of the feature vectors of each sample advertisement;

[0016] Obtain the first [item] through online search. Learning rate vector for each iteration Based on the first Base learner in round iteration In the Class learning rate Update to obtain the first Round-by-round iterative decision tree The middle belongs to the first Tree branches , ;

[0017] go through Iterative generation of decision tree The gradient boosting decision tree pre-training is complete.

[0018] Furthermore, links are established based on business relationships using a graph structure. The specific steps involved in building an advertising resource library using various resource sets are as follows:

[0019] The first Resource Collections by Category In Each advertisement feature vector is inserted into the feature space of the corresponding dimension. For class resource collection The first in The total number of advertisements is calculated using the kernel density estimation method based on the density distribution, and the centroid of the advertisement feature is used as the first centroid. Ad-like feature vectors ;

[0020] Self-supervised learning training is performed using the Transformer model to train the first... Ad-like feature vectors By mining dependency relationships through attention mechanisms, the first generation is generated. Class Reconstruction of Ad Feature Vectors ;

[0021] Sequentially obtain the first type of reconstructed advertisement feature vector To the Class Reconstruction of Ad Feature Vectors And as a graph node, initialize the edges between graph nodes to construct a fully connected graph, where each edge includes edge weight and edge direction;

[0022] A graph attention network is used to learn and determine the business relationships and upstream / downstream connections between graph nodes through an attention mechanism. The edges between the nodes in the graph complete the construction of the advertising resource library.

[0023] Specifically, the equipment installation information includes the equipment installation address and the name of the installation company. Business information of the target community is obtained through online searches and recorded in text form. Natural language processing technology is then used to transform this information into a feature vector of the target community. Analyzing the feature vectors of the target community using the BERT model The data is then categorized into corresponding resource sets. It is checked whether similar community ad interaction records already exist within those resource sets. If they do, a federation is constructed using a horizontal migration recommendation algorithm. Recommendations are then made to target communities based on the general interests of the federation's ad interaction records. If an ad does not exist, it will be randomly selected from the candidate ad set. The individual ads are recommended to the target community, while the candidate ad set is a collection of ads from the target community's business and its upstream and downstream businesses. This refers to the number of ads recommended in a single instance.

[0024] Furthermore, a lateral migration recommendation algorithm is used to recommend to the target community. An advertisement includes the following specific steps:

[0025] Assuming the target community feature vector Classified to No. Resource Collections by Category Searching the ad resource library for the first Resource Collections by Category The resource set of classes whose edge weights are greater than or equal to the edge weight threshold, and the first Resource Collections by Category Merge into a candidate ad set;

[0026] To acquire similar communities and build a federation, during the ad recall phase, a matrix factorization model is used to multiply the community matrix of each similar community with the ad feature vector in the candidate ad set to filter out the ads of interest for each similar community. The community matrix is ​​obtained by organizing the ad interaction records of the community.

[0027] In the ranking stage, a deep interest network model is selected. The interest extraction layer considers the temporal sequence of advertising interaction records in each similar community, and the attention weights are used to understand the real-time interest status of similar communities. Interested advertisements are scored and prioritized to generate an advertising interest sequence. And select the previous Recommend an advertisement. This refers to the number of ads recommended in a single instance.

[0028] Specifically, the matrix factorization model and the deep interest network model are deployed in each community of the same type in the federation for distributed training. Differential privacy technology is used in each community of the same type to protect the data security of the training results. The target community collects the model training results of the communities of the same type and calculates the average gradient of each model through the federated stochastic gradient descent method. The parameters are updated through stochastic gradient descent and fed back to the corresponding communities of the same type. Pre-training is completed by setting a loss function threshold.

[0029] Specifically, the water purification equipment uses a flow sensor to detect water intake actions, generating and storing water intake records. It also uses visual detection technology and QR code scanning technology to detect users' attention to and viewing of advertisements within the target community, generating and storing advertisement interaction records. Specifically, when an attention-based advertisement is detected, an advertisement interaction tag is generated. Setting it to 0 will detect ad interaction tags when viewing ads. Set to 1 to record the ad name and ad interaction tags of the ads that are noticed or viewed. and ad interaction time .

[0030] Furthermore, sensing user attention to advertising behavior through visual detection technology includes the following specific steps:

[0031] Face images were detected using the RetianFace face detection algorithm. The head roll angle is obtained separately based on the direction of the line segment connecting the inner corners of the eyes and the left and right corners of the mouth. ;

[0032] Based on head roll angle Calculate the rotation matrix to transform the face image. Perform image alignment and eliminate head roll angle. The impact of 2D Euler angle gaze angle estimation on obtaining aligned face images ;

[0033] Align face images Input a 2D Euler angle gaze estimation network and use ResNet50 to analyze and align face images. The texture information in the image is used to generate facial feature maps. ;

[0034] Introducing squeeze excitation operation to filter facial feature maps Irrelevant features in the image are integrated and segmented through a fully connected layer to generate a pitch-view feature map. and yaw line-of-sight feature map ;

[0035] The pitch and line-of-sight feature maps are predicted using the Softmax function. and yaw line-of-sight feature map The classification probability at each angle setting, and the pitch-view feature map. and yaw line-of-sight feature map The expected angle as the human eye's pitch angle at all angle settings and the yaw angle of the human eye ;

[0036] Based on head roll angle human eye pitch angle and the yaw angle of the human eye Derive and obtain the gaze direction vector and determine whether it intersects with the advertisement display area;

[0037] If there is no intersection, no action is taken; if there is an intersection, the ad interaction time is recorded. It continues to track until the gaze direction vector does not intersect with the advertisement display area, then counts the attention duration and compares it with the attention duration threshold;

[0038] If the attention duration is less than the threshold, delete the record of the ad interaction time. If the attention duration threshold is greater than or equal to the ad interaction tag, the ad interaction tag will be set. Setting it to 0, combined with the ad name and ad interaction tags of the ads being noticed. and ad interaction time Generate ad interaction records.

[0039] Furthermore, a squeeze-excitation operation is introduced to filter facial feature maps. The irrelevant features in the process include the following specific steps:

[0040] Facial feature map The two-dimensional component feature maps of each channel are aggregated into a scalar, establishing the connection between different channels to obtain the face feature vector. ;

[0041] The facial feature vector is processed through two fully connected layers. To enhance non-linear representation capabilities, proportional dimensionality reduction and expansion are performed, and the vector is normalized using the Sigmoid function to resemble a facial feature vector. Same-dimensional activation vector ;

[0042] The activation vector The excitation value of each channel is used as a facial feature map. The weights of the corresponding channels are multiplied together to obtain the facial feature map. Adjustments will be made.

[0043] Furthermore, after each adjustment cycle, the stored water collection records in the target community are statistically analyzed to obtain drinking water statistics. The drinking water statistics include the total water collection volume and the total number of water collections within a single adjustment cycle. The population size is calculated and similar communities are identified through fuzzy inference, including the following specific steps:

[0044] The total water intake and total number of water intakes of the target community are used as fuzzy inputs, and the population size of the target community is used as the fuzzy output. The first fuzzy set is defined based on the range of total water intake, including low, medium and high. The second fuzzy set is defined based on the range of total number of water intakes, including few, medium and many. The third fuzzy set is defined based on the range of population size, including small, relatively small, medium, relatively large and large.

[0045] The first fuzzy set, the second fuzzy set, and the third fuzzy set are described by triangular membership functions, and fuzzy rules are formulated to establish the mapping from the combination of the first fuzzy set and the second fuzzy set to the third fuzzy set.

[0046] Based on the membership function, calculate the membership degree of the total water intake with respect to each first fuzzy set and the membership degree of the total number of water intakes with respect to each second fuzzy set. Combine the non-zero membership degrees in the first and second fuzzy sets to activate the fuzzy rules and match them to the corresponding third fuzzy set. Select the smaller membership degree in the combination as the membership degree of the corresponding third fuzzy set. Select the third fuzzy set with the largest membership degree as the population size degree of the target community.

[0047] Select communities of the same type as the target community in terms of population size as similar communities.

[0048] Furthermore, recommendations for the target community are updated by fusing filtering recommendation algorithms. An advertisement includes the following specific steps:

[0049] Obtain the candidate ad set of the target community, and reconstruct the sub-federation of the target community and similar communities. Obtain the score of each ad in the candidate ad set through the horizontal migration recommendation algorithm and normalize it into interest probability. Arrange the ads in the candidate ad set according to their storage order to generate the first interest probability sequence.

[0050] Obtain the ad interaction records of the target community, based on the ad interaction time. Sort the ad feature vectors corresponding to the ad names and generate ad click sequences. ;

[0051] Using a random walk algorithm in the ad click sequence The sampling start point is randomly selected to sample and obtain the advertising interaction time. Nine consecutive ad feature vectors are used to construct an ad click sampling sequence; this step is repeated. Next, obtain A sequence of ad clicks This represents the total number of samples.

[0052] Using the Skip-Gram model, the position located at the th A sample sequence of ad clicks The feature vector of the middle ad is considered the center word. The feature vectors of the first four ads and the last four ads are considered the preceding and following text, respectively. The preceding and following text are predicted by multiplying the weight matrix with the center word, and the dimensionality is reduced to generate the next text. A dense click feature sequence , ;

[0053] Will Each dense click feature sequence is mapped to a different feature space through multiple self-attention mechanisms, and the feature correlation within each dense click feature sequence is calculated. The outputs of multiple self-attention mechanisms are then concatenated using a linear matrix to generate a new feature sequence. A reconstructed click feature sequence;

[0054] Will The reconstructed click feature sequence is adjusted in dimension by max pooling and linear matrix and then linearly concatenated with each ad feature vector in the candidate ad set. The interest probability of each ad feature vector is generated by activating the linear concatenation result through the Sigmoid function.

[0055] The interest probabilities are arranged according to the storage order of the advertisements in the candidate advertisement set to generate a second interest probability sequence. The first and second interest probability sequences are then weighted and summed using a pre-set weight ratio to generate a fused interest probability sequence. The interest probability sequence with the highest interest probability is selected from the fused interest probability sequence. Recommend an advertisement. This refers to the number of ads recommended in a single instance.

[0056] A media delivery system based on water purification equipment data analysis is used to execute the media delivery method based on water purification equipment data analysis, including a data acquisition and detection module, a storage and sharing module, an advertising recommendation module, and a display module;

[0057] The data collection and detection module senses water-taking actions and collects water-taking records from the target community. It also senses advertising interaction behavior and collects advertising interaction records through visual detection technology and QR code scanning detection technology.

[0058] The storage and sharing module stores the water collection records and advertising interaction records of the target community, and shares the advertising resource library through cloud technology;

[0059] When water purification equipment is installed, the advertising recommendation module obtains information text of the target community based on the equipment installation information and classifies it using a text classification algorithm. Based on the existence of advertising interaction records in similar communities, it decides whether to use a horizontal migration recommendation algorithm or to recommend advertisements through random sampling. After each adjustment cycle, it obtains drinking water statistics data and uses fuzzy reasoning to calculate the population size to determine similar communities. It then uses a fusion filtering recommendation algorithm to update the recommended advertisements by considering the advertising interaction records of the target community and similar communities.

[0060] The display module shows the content of the recommended advertisements.

[0061] Furthermore, the data acquisition and detection module includes a flow sensing unit, a visual detection unit, and a barcode scanning detection unit;

[0062] The flow sensing unit uses a flow sensor to detect water intake actions, collects and generates water intake records, and sends them to the storage and sharing module.

[0063] The visual inspection unit acquires facial images by capturing images through a visual inspection device. Visual detection technology was used for analysis, and the RetianFace face detection algorithm was used to identify the inner corners of the eyes and the left and right corners of the mouth to calculate the head roll angle. And align the face image A two-dimensional Euler angle gaze estimation network is input, ResNet50 is used to extract features and a squeeze excitation operation is introduced to filter irrelevant features. The probability distribution is predicted through fully connected layers and a Softmax function, and the human eye pitch angle is calculated. and the yaw angle of the human eye Generate a gaze direction vector and determine whether it intersects with the advertising display area. Based on the intersection, decide whether to generate an advertising interaction record and transmit it to the storage sharing module.

[0064] The built-in QR code scanning detection unit uses QR code scanning detection technology to sense scanning signals, automatically generate advertising interaction records, and transmit them to the storage and sharing module.

[0065] Furthermore, the advertising recommendation module includes a migration recommendation unit and a personalized recommendation unit;

[0066] During the installation of water purification equipment, the migration recommendation unit retrieves information text from the target community based on the equipment installation information. It then uses a text classification algorithm to categorize the text into corresponding resource sets and determines if there are advertising interaction records from similar communities. If so, a lateral migration recommendation algorithm is used to filter suitable advertisements based on the advertising interaction records from all similar communities, prioritizing them and selecting the highest-priority advertisement. The ad is recommended to the target community; if it does not exist, it is randomly selected from the candidate resource set. Recommend an advertisement. This represents the total number of recommendations made in a single ad session.

[0067] After each adjustment cycle, the personalized recommendation unit generates drinking water statistics and uses fuzzy inference to calculate the population size. It then filters out similar communities with the same population size from other similar communities. Finally, it analyzes the advertising interaction records of similar and target communities using a fusion filtering recommendation algorithm to reselect... Recommend an advertisement.

[0068] Compared with existing technologies, the significant advantages of this invention are as follows: It uses a gradient boosting decision tree to divide advertising resource sets into class resource sets based on business information, and links different class resource sets through a graph structure to reflect business relationships. Based on equipment installation information, it connects to the network to obtain information text from target communities and uses a text classification algorithm to categorize it into the corresponding class resource sets. Based on the existence of advertising interaction records in similar communities, it uses a lateral migration recommendation algorithm or random sampling to recommend business-specific advertisements to the target community. After collecting water collection records and advertising interaction records from the target community for a certain period, it calculates the population size and identifies similar communities through fuzzy reasoning. Finally, it uses a fusion filtering recommendation algorithm to consider the advertising interaction records of the target community and similar communities, providing targeted advertising recommendations for the target community. This greatly improves the accuracy and effectiveness of advertising placement, achieving an effective combination of water purification equipment data and advertising media. Attached Figure Description

[0069] Figure 1 A flowchart illustrating a media placement method based on data analysis from water purification equipment;

[0070] Figure 2 This is a flowchart of the visual inspection technology in this invention;

[0071] Figure 3 This is a schematic diagram of the fusion filtering recommendation algorithm in this invention;

[0072] Figure 4 This is an application scenario diagram of the media delivery system based on water purification equipment data analysis in this invention. Detailed Implementation

[0073] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0074] Example 1

[0075] like Figure 1 As shown, a specific embodiment of the present invention discloses a media placement method based on data analysis of water purification equipment, including the following specific steps:

[0076] Get advertising resource set Based on business information, a gradient boosting decision tree is used to divide the advertising resource set. To build Each type of resource set will be categorized based on business relationships. The various resource sets are linked in a graph structure to build an advertising resource library. The business information includes the business theme, target audience, and marketing objectives. Total number of advertising categories;

[0077] Based on device installation information, the system retrieves textual information about the target community from the internet and uses a text classification algorithm to categorize the target community into corresponding resource sets. The system then uses the presence or absence of advertising interaction records from similar communities to determine whether to use a lateral migration recommendation algorithm or random sampling to select and recommend the target community. In this advertisement, "community" is defined as a group of fixed and / or transient people with a certain base. In this embodiment, "community" specifically refers to the company or merchant that installs water purification equipment, and "target community" refers to the company or merchant that newly installs water purification equipment. This refers to the number of ads recommended in a single instance.

[0078] The system collects and stores water intake records and advertising interaction records of the target community in real time. It also compiles drinking water statistics within each adjustment cycle, establishes a mapping from drinking water statistics to population size through fuzzy inference, and further identifies similar communities. Finally, it uses a fusion filtering recommendation algorithm to consider advertising interaction records between the target community and similar communities in a weighted manner, and re-recommends recommendations for the target community. The advertisement mentions that similar communities refer to communities of the same type with the same population size.

[0079] Furthermore, advertising resource collection ,in, For the first The feature vector of the first advertisement can be processed using natural language processing techniques. The business information of each advertisement is analyzed and extracted, and further input is performed. Decision trees generated by gradient boosting decision tree algorithm in rounds of pre-training Based on the Softmax function, the first Each advertising feature vector In decision tree China regarding The scores of each class are transformed into a probability distribution to obtain the scores of the classes belonging to the first class. class probability The specific formula is as follows:

[0080] ,

[0081] in, For advertising feature vectors, For the first Each advertising feature vector Class tags, For the first Each advertising feature vector In decision tree Regarding the first Class score, number Each advertising feature vector Predictive tags The final choice is The class with the highest probability among all classes, i.e. , , , , and These represent the total number of ads, the total number of training rounds of the decision tree, and the total number of ad categories, respectively, after processing by the decision tree. And the processing of the Softmax function, advertising resource set Each ad feature vector in the dataset is assigned a corresponding class label, and the dataset is divided into three categories. Each type of resource set is denoted as... ,in, The class tag is the first The set of all advertising feature vectors of a class.

[0082] Furthermore, compared to traditional decision tree algorithms, gradient boosting decision trees gradually reduce the classification loss of the decision tree by training a new base learner in each iteration to fit the residual of the previous iteration, and then weighted summing. The base learners trained in the rounds of iterations produce the decision tree. , To determine the total number of training rounds for the decision tree, the base learners are classification and regression trees. The pre-training of the gradient boosting decision tree includes the following specific steps:

[0083] Get advertising sample set ,in, and The first Sample The sample advertisement feature vectors and corresponding sample class labels are used, where the sample class labels are pre-annotated manually to provide real values ​​for comparison during the pre-training of the gradient boosting decision tree. , This represents the total number of sample advertisements, where "sample" specifically refers to advertisements that have been labeled with sample category tags.

[0084] Initialize a constant predictor , No. Each sample advertisement feature vector After constant predictor The information obtained All classes have the same score, i.e., the score is the same. , , Total number of advertising categories;

[0085] In the In the round of iteration, , For the total number of training rounds of the decision tree, for the th Each sample advertisement feature vector In the Decision tree generated in rounds of iteration Classification loss in Solve the problem about the decision tree separately. middle The partial derivative of the score of each class is used to obtain the score of the first class. In the first iteration Each sample advertisement feature vector Belonging to The negative gradient of each class is given by the following formula:

[0086] ,

[0087] in, For the first In the first iteration Each sample advertisement feature vector Belongs to the The negative gradient of the class, For the first Each sample advertisement feature vector In decision tree Regarding the first The class score, due to the decision tree Used to solve multi-class classification problems, so classification loss The cross-entropy loss is as follows:

[0088] ,

[0089] in, This is an indicator function; when the condition within the parentheses is met, When the conditions within the parentheses are not met, , For the first Each sample advertisement feature vector In decision tree The middle belongs to the first The probability of the class, at which point the negative gradient... Specifically ;

[0090] Combination No. Each sample advertisement feature vector In the In each iteration, they belong to The negative gradient of the i-th class is used to construct the i-th class. Each sample advertisement feature vector In the Residual in round iteration , For the first Each sample advertisement feature vector In the In the round of iteration, belonging to the 1st The residuals of each class are used to train the first class. Base learner in round iteration Defit The residuals of the feature vectors of each sample advertisement minimize the fitting error, and the training and fitting of the base learner is the existing method.

[0091] Obtain the first [item] through online search. Learning rate vector for each iteration ,in, For the first Base learner in round iteration In the The learning rate for a class is calculated using the following formula:

[0092] ,

[0093] in, For the first Each sample advertisement feature vector Substitute the first Base learner in round iteration The first obtained from The fitted values ​​of the class, and based on the first Class learning rate Update to obtain the first Round-by-round iterative decision tree The middle belongs to the first Tree branches , For the first Round-by-round iterative decision tree The middle belongs to the first Tree branches of this type, Indicates the first Class learning rate A searchable set of learning rate values;

[0094] go through After rounds of iteration, the final decision tree is generated. The pre-training of the gradient boosting decision tree is complete.

[0095] Furthermore, based on business relationships... Each category of resource sets is linked in a graph structure. Building an advertising resource library includes the following specific steps:

[0096] For the Resource Collections by Category ,in, Representation class resource set The Middle Each advertising feature vector , For class resource collection The first in Total number of advertisements of each category;

[0097] Construct a feature space based on the dimensions of the advertising feature vector and then... The first ad feature vector is input, and the kernel density estimation method is used to calculate the ad feature centroid based on the density distribution of the ad feature vector in the feature space, which is then used as the first ad feature centroid. Resource Collections by Category The Ad-like feature vectors ;

[0098] Self-supervised learning training is performed using the Transformer model to train the first... Ad-like feature vectors Inputting the Transformer model, the representation is reconstructed through an attention mechanism to mine the th... Ad-like feature vectors The dependencies between them are studied, and the mutual influence between audience goals and marketing objectives in the reconstructed representation is learned to generate the first... Class Reconstruction of Ad Feature Vectors ;

[0099] Sequentially obtain the first type of reconstructed advertisement feature vector To the Class Reconstruction of Ad Feature Vectors And as a graph node, the edges between graph nodes are randomly initialized to construct a fully connected graph. The edges include edge weight and edge direction, which represent the association and upstream and downstream relationship between two connected graph nodes, respectively.

[0100] A graph attention network is used to learn and determine the business relationships and upstream / downstream connections between different graph nodes through an attention mechanism. The edges between the nodes of the graph complete the construction of the advertising resource library. The Transformer model and the graph attention network are pre-trained through a self-supervised task. The self-supervised task is generally based on random masking and sample comparison, which is a conventional technique and will not be described in detail.

[0101] Specifically, the equipment installation information includes the equipment installation address and the name of the installation company. Business information of the target community is obtained through online searches and recorded in text form, and then transformed into a feature vector of the target community using natural language processing technology. Based on the target community feature vector using the BERT model The data is categorized into corresponding resource sets. It is then checked whether similar community ad interaction records already exist within those resource sets. If such records do exist, a federation is constructed using a lateral migration recommendation algorithm. Based on the common interests of the federation's ad interaction records, recommendations are made to the target communities. If no similar community has a record of ad interaction for a given ad, then a candidate ad will be randomly selected from the candidate ad set. The system recommends ads to the target community, and the ad interaction records include ad name and ad interaction tags. and ad interaction time The candidate ad set is a collection of ads from the target community's business and its upstream and downstream businesses. This refers to the number of ads recommended in a single instance.

[0102] Furthermore, a lateral migration recommendation algorithm is used to recommend to the target community. An advertisement includes the following specific steps:

[0103] Assuming the target community feature vector Classified to No. Resource Collections by Category That is, the first Resource Collections by Category The advertisements within the platform align with the target community's business. Based on the edges between different resource sets within the advertisement resource library, the system searches for advertisements that match the target community's business. Resource Collections by Category The resource set of classes whose edge weights are greater than or equal to the edge weight threshold, and the first Resource Collections by Category Merge into a candidate ad set for the target community;

[0104] Since the target community has just installed water purification equipment, no effective drinking water data or advertising interaction records have been collected. Therefore, it is impossible to perform personalized advertising recommendations through the fusion filtering recommendation algorithm and obtain the data categorized as follows: Resource Collections by Category The remaining communities of the same type are federated. Since the federation is built from communities of the same type, the advertising interaction records of communities of the same type in the federation have common patterns and characteristics, which can temporarily replace the personalized interaction data of the target community.

[0105] During the ad recall phase, each community of the same type in the federation adopts a low-complexity matrix factorization model to organize the ad interaction records of each community into a community matrix. The community interests and ad features are modeled by multiplying the matrix with the ad feature vectors in the candidate ad set. The main purpose of the ad recall phase is to perform initial screening of the ads in the candidate ad set and select the ads of interest for each community of the same type.

[0106] During the ranking phase, each community of the same type in the federation uses a deep interest network model. Through an interest extraction layer, the interaction sequence in the ad interaction records is considered. By combining historical and recent ad interaction behaviors of the same type of community with attention weights, the real-time interest status of the same type of community is accurately understood. The ads of interest selected in the ad recall phase are scored, and they are prioritized from high to low according to the scores to generate an ad interest sequence. And select the previous Several advertisements were recommended, among which... This refers to the number of ads recommended in a single instance.

[0107] Specifically, the pre-training of the matrix factorization model and the deep interest network involves deploying the matrix factorization model and the deep interest network model separately in each similar community within the federation. Each similar community performs distributed training based on its own advertising interaction records, and differential privacy technology is used to add noise to the training results to ensure data security. The target community collects the model training results from each similar community and performs weighted averaging and gradient fusion on the model using the federated stochastic gradient descent method. That is, the learning rate is fixed and the average gradient of each model is calculated. The parameters are updated using stochastic gradient descent and further fed back to each similar community. Pre-training is completed by setting a loss function threshold. Here, differential privacy technology is a widely used data privacy protection mechanism and will not be described in detail.

[0108] Specifically, the water purification equipment uses a high-precision flow sensor to accurately detect water intake actions, generating and storing water intake records. These records include the water intake time, volume, and duration. Visual detection technology and QR code scanning technology are used to detect users' attention to and viewing of advertisements within the target community, generating and storing advertisement interaction records. Specifically, when a user's gaze is detected on an advertisement, an advertisement interaction tag is generated. Set to 0, and when a user scans a QR code to view an advertisement, the ad interaction tag will be set to 0. Set to 1, and record the ad name and ad interaction tags of the ads that are noticed or viewed. and ad interaction time .

[0109] like Figure 2 As shown, further, sensing user attention to advertisements through visual detection technology includes the following specific steps:

[0110] Obtain face image Since the direction of human vision is mainly determined by the head roll angle human eye pitch angle and the yaw angle of the human eye Decision, and head roll angle The change only causes the human eye's visual image to rotate within the same imaging plane. To reduce the excessive complexity of joint estimation of three angles in traditional visual detection algorithms, the RetianFace face detection algorithm is used to detect face images. Coordinates of the inner canthus of the left eye Coordinates of the inner canthus of the right eye Coordinates of the left corner of the mouth and the coordinates of the right corner of the mouth ,in, and The x and y axes are the horizontal and vertical axes of the image's two-dimensional coordinate system, respectively. The head roll angle is obtained separately by combining the direction of the line segment connecting the center of the inner corners of both eyes with the center of the line connecting the left and right corners of the mouth. The specific formula is as follows:

[0111] ;

[0112] in, Represents the inverse cotangent function;

[0113] Based on head roll angle For facial images Image alignment is performed by creating a virtual camera aligned with the center point of the face, and then using a rotation matrix to adjust the coordinate system of the virtual camera. Axis aligned to face image The direction of the line connecting the inner corners of both eyes is used to eliminate the head roll angle. The impact of 2D Euler angle gaze angle estimation on obtaining aligned face images ;

[0114] Align face images Input a 2D Euler angle gaze estimation network, use ResNet50 as the feature extractor, and analyze the aligned face image. The texture information in the image is used to generate facial feature maps. ResNet50 is an existing residual network series. The face feature map generated after processing by ResNet50 is... Dimensions ,in, The number of channels in ResNet50. and Facial feature maps The height and width are equal to the height of the face image. height With width The value obtained by subtracting the kernel size, adding the stride and twice the padding width, and then dividing by the stride;

[0115] Because ResNet50 directly obtains facial feature maps Face feature maps with high dimensionality and irrelevant features are filtered by introducing a squeeze excitation operation. Irrelevant features in the image are integrated and segmented using a fully connected layer to generate a pitch-view feature map. and yaw line-of-sight feature map ;

[0116] The pitch and line-of-sight feature maps are predicted using the Softmax function. and yaw line-of-sight feature map The classification probability is calculated for each angle setting, and the pitch and line-of-sight feature maps are obtained separately. and yaw line-of-sight feature map The desired angle of human eye tilt at all angle settings and the yaw angle of the human eye Among them, the angle settings are divided into intervals of 2 degrees;

[0117] Based on head roll angle human eye pitch angle and the yaw angle of the human eye The gaze direction vector is derived, and spatial geometry is used to determine whether there is an intersection between the gaze direction vector and the advertising display area.

[0118] If there is no overlap, it is determined that the user did not notice the advertisement and no action is taken; if there is overlap, the advertisement interaction time is recorded. It continuously tracks the user's gaze direction vector until the gaze direction vector does not intersect with the advertisement display area, then counts the attention duration and determines whether the attention duration is greater than or equal to the attention duration threshold.

[0119] If the attention duration is less than the threshold, it is considered invalid attention, and the recorded ad interaction time is deleted. ;

[0120] If the attention duration is greater than or equal to the attention duration threshold, it is considered valid attention, and the ad interaction tag will be added. Set to 0, and further specify the ad name and ad interaction tags for the ads to be noticed. and ad interaction time Integrate and store the generated advertising interaction records.

[0121] Furthermore, a squeeze-excitation operation is introduced to filter facial feature maps. The irrelevant features in the process include the following specific steps:

[0122] Based on facial feature maps width and height Perform a squeezing operation to obtain the facial feature vector. The squeezing operation is essentially the process of compressing facial feature images. The two-dimensional component feature maps of each channel are integrated into a scalar to establish the connection between different channels. The specific formula is as follows:

[0123] ,

[0124] in, facial feature vector The Middle Scalar values ​​for each channel, For facial feature maps No. The height of each channel is And the width is Feature values ​​at the location, Therefore, facial feature vectors The dimension is , The number of channels in ResNet50;

[0125] The facial feature vector is processed through two consecutive fully connected layers. Dimensionality reduction and expansion are performed proportionally, while maintaining the same dimensionality. Nonlinear representation capabilities are added through nonlinear functions in the fully connected layer, and further scaled to the range of 0 to 1 using the Sigmoid function, transforming the vector to generate a facial feature vector. Same-dimensional activation vector , activation vector This indicates the importance of each channel;

[0126] Through the activation vector facial feature map Recalibrate, that is, recalibrate the excitation vector. The excitation value of each channel is used as a facial feature map. The weights of the corresponding channels are multiplied together to assign a weight to the facial feature map based on the importance of each channel. Adjustments were made so that the features of high-importance channels were largely preserved, while the weights of low-importance channels were smaller, and the product of their corresponding features was almost zero.

[0127] Furthermore, after each adjustment cycle, the stored water collection records in the target community are statistically analyzed to obtain drinking water statistics. The drinking water statistics include the total water collection volume and the total number of water collections within a single adjustment cycle. Fuzzy reasoning is used to establish a mapping between drinking water statistics and population size, and similar communities are further identified, including the following specific steps:

[0128] Using the total water withdrawal amount and total number of withdrawals of the target community as fuzzy inputs, and the population size of the target community as the fuzzy output, we define the following fuzzy sets: a first set corresponding to the total water withdrawal amount, including low (less than or equal to 2000 liters), medium (greater than 2000 liters and less than or equal to 5000 liters), and high (greater than 5000 liters); a second set corresponding to the total number of withdrawals, including few (less than or equal to 200 times), medium (greater than 200 times and less than or equal to 500 times), and many (greater than 500 times); and a third set corresponding to the population size, including small (less than or equal to 50 people), relatively small (greater than 50 people and less than or equal to 200 people), medium (greater than 200 people and less than or equal to 500 people), relatively large (greater than 500 people and less than or equal to 1000 people), and large (greater than 1000 people).

[0129] Triangular membership functions are used to describe the first, second, and third fuzzy sets, respectively. Further fuzzy rules are then formulated. These rules serve as guidelines for the inference algorithm during fuzzy inference and are used to establish the mapping from the combination of the first and second fuzzy sets to the third fuzzy set. Since there are three sets each for the first and second fuzzy sets, a total of nine fuzzy rules are established, as follows:

[0130] If the total amount of water taken is low and the total number of water take-offs is infrequent, then the scale of personnel is small.

[0131] If the total amount of water taken is low and the total number of water taken is medium, then the scale of personnel is relatively small.

[0132] If the total amount of water taken is low but the total number of water take-offs is high, then the population size is medium.

[0133] If the total water intake is medium and the total number of water intakes is low, then the scale of personnel is relatively small.

[0134] If the total water intake is medium and the total number of water intakes is medium, then the personnel scale is medium.

[0135] If the total amount of water taken is medium and the total number of water takings is high, then the scale of personnel is relatively large.

[0136] If the total water intake is high and the total number of water intakes is low, then the population size is medium.

[0137] If the total water intake is high and the total number of water intakes is medium, then the scale of personnel is relatively large.

[0138] If the total amount of water taken is high and the total number of water take-offs is high, then the scale of personnel is large.

[0139] Based on the membership function, calculate the membership degree of the total water intake with respect to each first fuzzy set and the membership degree of the total number of water intakes with respect to each second fuzzy set. Obtain the membership degrees that are not zero in the first and second fuzzy sets, combine and activate the corresponding fuzzy rules and match them to the corresponding third fuzzy set. Select the smaller membership degree in the combination as the membership degree of the corresponding third fuzzy set.

[0140] The membership degrees of the third fuzzy set obtained from all activated fuzzy rules are combined by taking the largest value, and the third fuzzy set with the largest membership degree is selected as the population size of the target community.

[0141] From all communities of the same type that correspond to the target community with the same resource set, further select communities of the same type with the same population size as the target community, and designate them as similar communities.

[0142] like Figure 3 As shown, further, recommendations for the target community are updated by fusion filtering recommendation algorithms. An advertisement includes the following specific steps:

[0143] Candidate ad sets for the target community are obtained based on the edges between different class resource sets in the ad resource library;

[0144] The target community and similar communities are reconstructed into sub-federations. The score of each advertisement in the candidate advertisement set is obtained through the horizontal migration recommendation algorithm and normalized to become the interest probability. The first interest probability sequence is generated by arranging the advertisements in the candidate advertisement set according to their storage order.

[0145] Obtain the ad interaction records of the target community, based on the ad interaction time. The ad click sequence is generated by sorting the ad feature vectors corresponding to the ad names in order. ;

[0146] Due to ad click sequence The time span is a single adjustment period, which may involve changes in the target community's interests. For example, the target community's interests may shift from upstream business-related ads to downstream business-related ads and then back to upstream business-related ads within a single adjustment period. To better capture the target community's short-term interests, a random walk algorithm is used in the ad click sequence. The algorithm randomly selects an ad feature vector as the sampling starting point and moves through the sampled vectors to obtain the data during ad interaction time. Use nine consecutive ad feature vectors to construct an ad click sampling sequence, and repeat this step. Next, obtain The ad click sampling sequences are denoted as follows: ,in, For the first The ad click sampling sequence obtained from the second sampling. , This represents the total number of samples.

[0147] To improve recommendation efficiency, the Skip-Gram model is used to select ad click sampling sequences. The feature vector of the middle ad is considered the center word. The feature vectors of the first four ads and the last four ads are considered the context and background of the center word, respectively. The context and background are predicted by multiplying the weight matrix with the center word, and the dimensionality is reduced to generate the next word. A dense click feature sequence , ;

[0148] A dense click feature sequence They respectively represent The feature associations between advertising feature vectors within different short-term time periods can effectively reflect the interests of the target community within those short-term time periods, and further... A dense click feature sequence A multi-head attention mechanism consisting of four self-attention mechanisms is used to map the data to different feature spaces, and the feature correlation within each dense click feature sequence is calculated separately. The outputs of the four self-attention mechanisms are then concatenated using a linear matrix to generate the final feature. A reconstructed click feature sequence , For the first A dense click feature sequence The reconstructed click feature sequence generated after processing is comparable to the dense click feature sequence. Reconstructing the click feature sequence More focus on key features;

[0149] For each ad feature vector in the candidate ad set, A reconstructed click feature sequence The interest probability of each ad feature vector is generated by adjusting the linear matrix to the same dimension as the ad feature vector through max pooling and linear concatenation with each ad feature vector. The linear concatenation result is then activated by the Sigmoid function.

[0150] The interest probabilities are arranged according to the ad storage order in the candidate ad set to generate a second interest probability sequence. Considering the influence of ad interaction records between the sub-federation and the target community on ad recommendation, the weight ratio of the sub-federation and the target community is set to 3:7. The first and second interest probability sequences are combined by weighted summation to generate a fused interest probability sequence. The interest probability with the highest interest probability in the fused interest probability sequence is selected. Recommend an advertisement. This refers to the number of ads recommended in a single instance.

[0151] Example 2

[0152] like Figure 4 In the application scenario shown, a specific embodiment of the present invention discloses a media delivery system based on water purification equipment data analysis, used to execute the media delivery method based on water purification equipment data analysis, including a data acquisition and detection module, a storage and sharing module, an advertising recommendation module, and a display module;

[0153] The data acquisition and detection module uses sensors to detect water-taking actions and collect water-taking records from the target community. It also uses visual detection technology and QR code scanning detection technology to detect advertising interaction behavior and collect advertising interaction records. Advertising interaction behavior includes noticing advertisements and viewing advertisements.

[0154] The storage sharing module stores the water collection records and advertising interaction records of the target community locally, and shares the advertising resource library with other communities through cloud technology;

[0155] When the device is first installed, the advertising recommendation module obtains information text of the target community based on the device installation information and classifies it using a text classification algorithm. Based on the existence of advertising interaction records in similar communities, it decides whether to recommend advertisements using a horizontal migration recommendation algorithm or by random sampling. After each adjustment cycle, it obtains drinking water statistics data, calculates the population size through fuzzy reasoning and identifies similar communities, and updates the recommended advertisements by considering the advertising interaction records of the target community and similar communities through a fusion filtering recommendation algorithm.

[0156] The display module shows the recommended advertisement content through the display device. The advertisement content includes the advertisement name, advertisement description, advertisement image and advertisement QR code.

[0157] Furthermore, the data acquisition and detection module includes a flow sensing unit, a visual detection unit, and a barcode scanning detection unit;

[0158] The flow sensing unit uses a high-precision flow sensor to detect water intake actions, collect and generate water intake records, and send them to the storage and sharing module.

[0159] The visual inspection unit acquires facial images by capturing images through a visual inspection device. Visual detection technology was used for analysis, and the RetianFace face detection algorithm was used to identify facial images. The head roll angle is calculated by using the inner corners of both eyes and the corners of the mouth. And align the face image The input is further processed by a two-dimensional Euler angle gaze estimation network. ResNet50 is used for feature extraction, and a squeeze excitation operation is introduced to filter irrelevant features. The human eye pitch angle is predicted through fully connected layers and a softmax function. and the yaw angle of the human eye The probability distribution is used, and the corresponding expectation is taken as the human eye pitch angle. and the yaw angle of the human eye The system obtains the gaze direction vector and, based on the intersection of the gaze direction vector with the advertising display area, decides whether to generate an advertising interaction record and transmit it to the storage sharing module.

[0160] The QR code detection unit has a built-in scanning signal receiving device and uses QR code scanning detection technology. When an external device scans the QR code to view an advertisement, the scanning signal receiving device senses the scanning signal, automatically generates an advertisement interaction record, and transmits it to the storage and sharing module.

[0161] Furthermore, the advertising recommendation module includes a migration recommendation unit and a personalized recommendation unit;

[0162] When the water purification equipment is initially installed, the migration recommendation unit retrieves information text from the target community based on the equipment installation information. It then uses a text classification algorithm to categorize the text into corresponding resource sets and determines if there are advertising interaction records from similar communities. If so, a lateral migration recommendation algorithm is used to filter suitable advertisements based on the advertising interaction records from all similar communities, prioritizing them and selecting the highest priority advertisement. Each ad is recommended to the target community. This represents the total number of recommendations made in a single ad session.

[0163] After the water purification equipment is deployed in the target community, the personalized recommendation unit queries the shared storage module to generate drinking water statistics data at each adjustment cycle. It then uses fuzzy inference combined with the drinking water statistics data to calculate the population size and further filters out similar communities with the same population size from the same type of community. Finally, it analyzes the advertising interaction records of similar communities and the target community using a fusion filtering recommendation algorithm to reselect... Recommend an advertisement.

[0164] This invention discloses a media placement method and system based on water purification equipment data analysis. It uses a gradient boosting decision tree to divide advertising resources into class resource sets based on business information, and links different class resource sets through a graph structure to reflect business relationships. Based on equipment installation information, it connects to the network to obtain information text from target communities and uses a text classification algorithm to categorize it into the corresponding class resource sets. Based on the existence of advertising interaction records in similar communities, it uses a lateral migration recommendation algorithm or random sampling to recommend business-specific advertisements to the target community. After collecting water collection records and advertising interaction records from the target community for a certain period, it estimates the population size and identifies similar communities through fuzzy reasoning. Finally, it uses a fusion filtering recommendation algorithm to consider the advertising interaction records of the target community and similar communities, providing targeted advertising recommendations for the target community. This significantly improves the accuracy and effectiveness of advertising placement, achieving an effective combination of water purification equipment data and advertising media.

[0165] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A media placement method based on data analysis of water purification equipment, characterized in that, The specific steps include the following: The advertising resource set is divided into multiple class resource sets based on business information by using a gradient boosting decision tree. All class resource sets are linked according to business relationships to build an advertising resource library. Information text of the target community is obtained based on device installation information and classified into the corresponding class resource set using a text classification algorithm. The existence of advertising interaction records in the same type of community determines whether to use a horizontal migration recommendation algorithm or random sampling to select recommended advertisements for the target community. During each adjustment cycle, water collection records are statistically analyzed to obtain drinking water statistics. The population size is calculated using fuzzy inference to determine similar communities. A fusion filtering recommendation algorithm is used to consider the advertising interaction records between the target community and similar communities, and advertisements are re-recommended for the target community. The method of recommending advertisements to a target community using the lateral migration recommendation algorithm includes the following steps: First, identify the resource set corresponding to the target community, and search the advertisement resource library for resource sets whose edge weights are greater than or equal to the edge weight threshold. Then, generate a candidate advertisement set from the resource set corresponding to the target community. Next, acquire similar communities, and use a matrix factorization model to multiply the community matrix of each similar community with the advertisement feature vector in the candidate advertisement set to filter out the advertisements of interest for each similar community. Finally, select a deep interest network model, consider the temporal sequence of advertisement interaction records for each similar community through an interest extraction layer, understand the real-time interest state of the similar community through attention weights, score the advertisements of interest and prioritize them, and select the advertisement with the highest priority. Recommend an advertisement. This refers to the number of ads recommended in a single instance. The fusion filtering recommendation algorithm for re-recommending ads to the target community includes the following steps: obtaining a candidate ad set for the target community; obtaining the score of each ad in the candidate ad set using a lateral migration recommendation algorithm and normalizing it into an interest probability; arranging the ads according to their storage order to generate a first interest probability sequence; obtaining ad interaction records for the target community; sorting the ad feature vectors corresponding to the ad names according to the ad interaction time; randomly selecting a sampling starting point using a random walk algorithm; sampling and obtaining multiple ad feature vectors with consecutive ad interaction times to construct an ad click sampling sequence; repeating this step to obtain multiple ad click sampling sequences; and using a Skip-Gram model to connect the weight matrix with each ad click sampling sequence. The product of the column's central words predicts the corresponding context, and dimensionality reduction generates the corresponding dense click feature sequence. Each dense click feature sequence is processed using a multi-head self-attention mechanism to calculate feature relevance in different feature spaces, and then linearly concatenated to generate the corresponding reconstructed click feature sequence. The dimension of each reconstructed click feature sequence is adjusted and concatenated with each ad feature vector in the candidate ad set. The interest probability of each ad feature vector is generated by activating the Sigmoid function. The interest probabilities are arranged according to the ad storage order in the candidate ad set to generate a second interest probability sequence. The first and second interest probability sequences are weighted and summed using weighted ratios to generate a fused interest probability sequence, and the sequence with the highest interest probability is selected. Recommend an advertisement.

2. The media placement method based on water purification equipment data analysis as described in claim 1, characterized in that, The advertising resource set includes Each ad feature vector is analyzed using natural language processing techniques to obtain business information about the ad. The gradient boosting decision tree algorithm uses the softmax function to optimize the ad feature vectors within the decision tree determined during pre-training. The scores of each class are transformed into probability distributions. Based on the class with the highest probability, an ad feature vector is assigned a predicted class label. The gradient boosting decision tree algorithm then divides the ad resource set into categories based on these predicted class labels. Individual resource sets, and These represent the total number of advertisements and the total number of advertisement categories, respectively.

3. The media placement method based on water purification equipment data analysis as described in claim 2, characterized in that, The pre-training of the gradient boosting decision tree algorithm includes the following specific steps: Obtain an ad sample set, which includes... Each sample advertisement feature vector and its corresponding sample class label, The total number of sample advertisements; Initialize the constant predictor; for each sample ad feature vector, generate a constant predictor with respect to... The scores of all classes are the same. Total number of advertising categories; In the In the round of iteration, , For the total number of training rounds of the decision tree, the feature vector of a single sample advertisement is trained on the th... Decision tree generated in rounds of iteration The classification loss in the decision tree is solved separately. middle The partial derivatives of the scores for each class are taken, and their negative values ​​are used as the corresponding negative gradients, where... For advertising feature vectors; Combining single-sample advertisement feature vectors in the th In each iteration, they belong to The negative gradients of each class are used to construct the feature vector of a single sample advertisement in the th case. The residual in the first iteration, training the second iteration Base learners in rounds of iteration to fit The residuals of the feature vectors of each sample advertisement; Obtain the first [item] through online search. The learning rate vector for each iteration, based on the first iteration. The base learner in the round of iteration is in the 1st round. The learning rate of the class is updated to obtain the first Round-by-round iterative decision tree The middle belongs to the first Tree branches , ; go through Decision tree generated after rounds of iteration .

4. The media delivery method based on water purification equipment data analysis as described in claim 1, characterized in that, The advertising interaction record is generated after sensing the user's advertising interaction behavior through visual detection technology and QR code scanning detection technology. The visual detection technology includes the following specific steps: Acquire facial images and identify the inner corners of the eyes and the left and right corners of the mouth using the RetianFace face detection algorithm. Obtain the head roll angle based on the line segment direction of the center of the line connecting the inner corners of the eyes and the center of the line connecting the left and right corners of the mouth. The face image is rotated based on the head roll angle to generate an aligned face image, which is then input into a two-dimensional Euler angle gaze angle estimation network. ResNet50 is used to analyze the texture information in the aligned face image to generate a face feature map. A squeeze excitation operation is introduced to filter irrelevant features in the face feature map, and a fully connected layer is used to integrate and segment to generate pitch and yaw line feature maps. The classification probabilities of the pitch and yaw line-of-sight feature maps at each angle level are predicted using the Softmax function, and the expected values ​​of all angle levels are calculated as the human eye pitch and yaw angles. The gaze direction vector is derived based on the head roll angle, eye pitch angle, and eye yaw angle, and it is determined whether it intersects with the advertising display area. If there is no intersection, no action is taken. If there is an intersection, the ad interaction time is recorded and continuously tracked until the gaze direction vector and the ad display area no longer intersect. Then, the attention duration is calculated and compared with the attention duration threshold. If the attention duration is less than the threshold, delete the ad interaction time. If it is greater than or equal to the attention duration threshold, set an ad interaction tag and combine the ad name, ad interaction tag, and ad interaction time to generate an ad interaction record.

5. The media placement method based on water purification equipment data analysis as described in claim 4, characterized in that, The process of introducing a squeeze excitation operation to filter irrelevant features in the face feature map includes the following specific steps: The two-dimensional component feature maps of each channel in the face feature map are aggregated into a scalar, and the connection between different channels is established to obtain the face feature vector. The face feature vector is reduced and increased in dimensionality proportionally by two fully connected layers, and then normalized to an activation vector by the Sigmoid function. The activation value of each channel in the activation vector is used as the weight value of the corresponding channel in the face feature map and multiplied together to adjust the face feature map.

6. The media delivery method based on water purification equipment data analysis as described in claim 1, characterized in that, The process of building an advertising resource library by linking all types of resource sets based on business relationships includes the following specific steps: All ad feature vectors in the resource set corresponding to a single class are placed into the feature space. The centroid of the ad feature is calculated based on the density distribution using the kernel density estimation method and used as the ad feature vector of the corresponding class. Self-supervised learning training is performed using the Transformer model. The dependency relationships of the corresponding class of advertising feature vectors are mined through the attention mechanism to generate the corresponding class of reconstructed advertising feature vectors. Obtain in sequence Reconstruct the advertising feature vectors of each class and use them as graph nodes. Initialize the edges between the graph nodes to construct a fully connected graph. Total number of advertising categories; By employing a graph attention network, the business relationships and upstream and downstream connections between graph nodes are learned through the attention mechanism, and the edges between each graph node are determined to complete the construction of the advertising resource library.

7. A media delivery system based on water purification equipment data analysis, used to execute the media delivery method based on water purification equipment data analysis as described in any one of claims 1-6, characterized in that, It includes a data collection and detection module, a storage sharing module, and an advertising recommendation module; The data collection and detection module senses water-taking actions and collects water-taking records of the target community. It also senses advertising interaction behavior and collects advertising interaction records through visual detection technology and QR code scanning detection technology. The storage and sharing module stores the water collection records and advertising interaction records of the target community, and shares the advertising resource library through cloud technology; The advertising recommendation module obtains information text of the target community based on the equipment installation information during water purification equipment installation, and classifies it using a text classification algorithm. Based on the existence of advertising interaction records in similar communities, it decides whether to use a horizontal migration recommendation algorithm or random sampling for advertising recommendation. After each adjustment cycle, it obtains drinking water statistics data and calculates the population size through fuzzy inference to determine similar communities. It then uses a fusion filtering recommendation algorithm to consider the advertising interaction records of the target community and similar communities and updates the recommended advertisements.

8. The media delivery system based on water purification equipment data analysis as described in claim 7, characterized in that, The advertising recommendation module includes a migration recommendation unit and a personalized recommendation unit; During the installation of the water purification equipment, the migration recommendation unit obtains information text from the target community based on the equipment installation information. It then uses a text classification algorithm to categorize the text into corresponding resource sets, determines whether there are advertising interaction records from similar communities, and if so, uses a lateral migration recommendation algorithm to filter suitable advertisements based on the advertising interaction records from all similar communities, prioritizing and selecting the advertisement with the highest priority. If an ad is not found, it will be randomly selected from the candidate resource set. Recommend an advertisement. This refers to the number of ads recommended in a single instance. Each time the personalized recommendation unit completes an adjustment cycle, it generates drinking water statistics and uses fuzzy inference to calculate the population size. It then filters out similar communities with the same population size from other similar communities, analyzes the advertising interaction records of similar and target communities using a fusion filtering recommendation algorithm, and reselects communities accordingly. Recommend an advertisement.

Images