Intelligent advertisement optimized delivery system based on user behavior analysis

Abstract

The invention discloses an intelligent advertisement optimized delivery system based on user behavior analysis, and the system comprises a data collection and preprocessing module which is used for collecting user behavior data, and carrying out the cleaning, analysis and processing of the user behavior data, so as to obtain the effective behavior data of a user; the user portrait construction module is used for constructing a user portrait by utilizing the effective behavior data of the user; the advertisement feature extraction module is used for performing multi-modal feature extraction on the advertisement materials to construct semantic representation of the advertisement; the conversion estimation model construction module is used for constructing a conversion estimation model through a training deep learning algorithm by utilizing the effective behavior data of the user and the advertisement features; the intelligent delivery decision module is used for constructing an intelligent delivery decision model; and an effect verification and optimization module. Through the scheme, user interests and advertisement contents can be accurately matched, advertisement sorting is optimized, the click rate and the conversion rate of advertisements are improved, meanwhile, invalid advertisement exposure is reduced, and the rate of return on investment of advertisers is improved.

Description

Technical Field

[0001] The present invention generally relates to the field of artificial intelligence technology. More specifically, the present invention relates to an intelligent advertising optimization delivery system and storage medium based on user behavior analysis. Background Art

[0002] In today's digital age, the Internet advertising industry is developing rapidly. However, with the explosive growth of Internet advertising, some pain points have gradually emerged.

[0003] On the one hand, users are often disturbed by a large number of advertisements when browsing the web or using mobile applications, which affects the browsing experience and causes users to resist advertisements, thereby affecting the click-through rate and conversion rate of advertisements, and ultimately affecting the effectiveness of advertising.

[0004] On the other hand, due to problems with the ad content or landing page design, the ad cannot achieve the expected coverage and reach, which may result in the ad failing to attract the user's interest and attention, resulting in a low click-through rate. Furthermore, even if the user clicks on the ad, the percentage of actual completed conversions (such as purchases, registrations, or downloads) is also low. This may be because the ad content does not match user needs, or the ad delivery strategy is not precise enough.

[0005] In view of the above problems, advertisers lack effective monitoring and optimization mechanisms during the advertising process, and are unable to find problems in time and make adjustments and improvements. This results in advertisers being unable to adjust strategies based on real-time data, thus missing out on opportunities to optimize advertising effects. As a result, the increase in advertising costs has increased the marketing burden on companies and affected commercial benefits.

[0006] In view of this, there is an urgent need to provide an intelligent advertising optimization delivery system based on user behavior analysis in order to improve the conversion effect of advertising. Summary of the invention

[0007] In order to at least solve one or more of the technical problems mentioned above, the present invention proposes an intelligent advertising optimization delivery system based on behavioral analysis in multiple aspects.

[0008] In a first aspect, the present invention provides an intelligent advertising optimization delivery system based on behavioral analysis. It includes:

[0009] Data collection and preprocessing module: used to collect user behavior data, and clean, analyze and process the user behavior data to obtain effective user behavior data; wherein the user behavior data includes browsing history, search history, and purchase behavior.

[0010] User portrait construction module: The configuration includes a user sentiment analysis unit and a user behavior pattern analysis unit; it is used to use the user's effective behavior data to construct a user portrait; the user portrait includes the user's attributes, interest preferences, and consumption habits; wherein the user sentiment analysis unit is configured to use the user's effective behavior data to analyze the user's speech and emotions on social media to obtain the user's emotional response to the advertisement. The user behavior pattern analysis unit is configured to use the user's effective behavior data to identify the user type, and the classification dimensions of the user type include the consumption time dimension, the consumption frequency dimension, and the consumption amount dimension.

[0011] Advertisement feature extraction module: used to perform multimodal feature extraction on advertisement materials to construct a semantic representation of the advertisement; wherein the multimodal feature extraction includes: extracting visual features using image recognition technology, and / or extracting text features using natural language processing technology.

[0012] Conversion prediction model building module: used to utilize the user effective behavior data and advertising features to build a conversion prediction model by training a deep learning algorithm to obtain the advertising conversion evaluation effect; the advertising conversion evaluation effect includes the click-through rate, bid, and relevance factors of the advertisement.

[0013] Intelligent delivery decision module: used to build an intelligent delivery decision model based on the user portrait, the semantic representation of the advertisement, and the advertisement conversion evaluation effect to generate an advertisement delivery strategy. This module includes: Intelligent sorting and matching unit: used to intelligently sort and match candidate advertisements based on the advertisement conversion evaluation effect using the LTP sorting model.

[0014] Effect verification and optimization module: configuration includes verification unit and delivery optimization unit; wherein, verification unit: configured to evaluate the delivery effects of different advertising strategies in real time through A / B test online experiment mechanism to verify the estimated effect of the advertising delivery strategy. Delivery optimization unit: configured to optimize the advertising sorting and matching according to the delivery effect feedback of the verification unit to improve the intelligent delivery decision model.

[0015] In some embodiments, using the user effective behavior data to analyze the user's speech and emotions on social media to obtain the user's emotional response to the advertisement includes: constructing an emotional vocabulary to perform sentiment analysis on the text; using a bag-of-words model, a TF-IDF model and / or a word embedding model to extract features from the text data; using a naive Bayes model to analyze the user's emotional tendency, the formula includes:

[0016]

[0017] Among them, P(C i |f1 ,f 2 ,...,f n ) is a given feature f 1 ,f 2 ,...,f n Conditional emotion category C i The probability, P(C i ) is the emotion category C i The prior probability, P(f i |C i ) is a given sentiment category C i Conditional feature f i The conditional probability of .

[0018] In some embodiments, the user type is identified by using the user effective behavior data, the steps include:

[0019] Step S4.1: extracting user behavior characteristics, including at least consumption time, consumption frequency, and consumption amount;

[0020] Step S4.2: Use K-means++ algorithm to select the initial cluster center;

[0021] Step S4.3: Use K-means algorithm to cluster user behavior patterns. The formula includes:

[0022]

[0023] Among them, J(C,μ) is the clustering quality function, which is used to evaluate the clustering effect; K is the number of clusters; C k represents the set of all data points in the kth cluster; μ k is the center point of the kth cluster; x is the data point in the cluster; ||||x-μ k |||| 2 is the square of the Euclidean distance between the data point x and the cluster center;

[0024] Step S4.4: constructing a multi-layer perceptron neural network for learning complex patterns of user behavior, wherein the multi-layer perceptron neural network comprises at least two hidden layers;

[0025] Step S4.5: Fuzzify the output of the K-means algorithm and the prediction result of the multi-layer perceptron neural network and convert them into fuzzy sets; and

[0026] Using the fuzzy set, a fuzzy rule base is constructed to explain the relationship between the user behavior pattern and the prediction result of the multi-layer perceptron neural network;

[0027] Step S4.6: The K-means clustering results and the multi-layer perceptron neural network prediction results are weighted fused. The formula includes:

[0028] FinalBehavior=α·K-means output+(1-α)NeuralNetwork Prediction,

[0029] Among them, α is a weight parameter with a value range of 0 to 1; K-means Output is the output of the K-means algorithm, indicating the cluster label to which each user data point belongs; Neural Network Prediction is the output of the multilayer perceptron neural network model, indicating the prediction result of the multilayer perceptron neural network on user behavior;

[0030] Step S4.7: using the fuzzy rule base to adjust the weights of different user behavior patterns to obtain clustering evaluation results; including using the silhouette coefficient indicator to evaluate the clustering effect;

[0031] Step S4.8: According to the clustering evaluation results, adjust the model parameters and fuzzy rules to optimize the type classification effect of the user behavior pattern.

[0032] In some embodiments, the deep learning algorithm includes: using a fully connected network to learn the relationship between user behavior data and advertising features, the formula includes:

[0033] P = FCN (F image ,F text ,U),

[0034] Among them, U is the user feature, P is the predicted conversion efficiency; F image is the extracted image feature, F text is the extracted text feature, and FCN is the fully connected network function.

[0035] In some embodiments, the delivery optimization unit includes: using a gradient descent method to update model parameters to optimize the advertisement ranking, and the formula includes:

[0036] θ t+1 =θ t -η▽ θ J(θ t ), where θ t represents the value of the model parameter at the tth iteration, θ t+1 represents the updated value of the model parameters at the t+1th iteration, η is the learning rate, which is used to control the step size of each step when updating the parameters, ▽ θ J(θ t ) is the gradient of the loss function J with respect to the parameter θ; J(θt ) represents the loss function at the tth iteration.

[0037] In a second aspect, the present invention provides a computer-readable storage medium, characterized in that a computer program is stored on the storage medium, and when the computer program is executed by a processor, the intelligent advertising optimization delivery system as described in any one of the first aspects is implemented.

[0038] Through the intelligent advertising optimization delivery system based on user behavior analysis as provided above, the embodiments of the present invention can train deep learning algorithms, build conversion prediction models, and obtain effective evaluations of advertising conversion effects by performing sentiment analysis and behavior pattern analysis on users, extracting visual features through graphic recognition and extracting text features through natural language processing on the multimodal features of advertising materials. In some embodiments, based on intelligent sorting and matching algorithms, an intelligent delivery decision model can be constructed, and the delivery strategy of the model can be optimized through feedback from actual delivery effects. Therefore, the intelligent delivery decision model can improve the relevance and acceptance of advertisements, reduce harassment of users, and improve the precise delivery effect of advertisements, thereby improving the conversion rate of advertisements and protecting the economic interests of advertisers. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] By reading the following detailed description with reference to the accompanying drawings, the above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily understood. In the accompanying drawings, several embodiments of the present invention are shown in an exemplary and non-limiting manner, and the same or corresponding reference numerals represent the same or corresponding parts, wherein:

[0040] Figure 1 An exemplary structural block diagram of an intelligent advertising optimization delivery system based on user behavior analysis according to an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0041] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0042] It should be understood that when the terms "first", "second", "third" and "fourth" are used in the claims, description and drawings of the present invention, they are only used to distinguish different objects, rather than to describe a specific order. The terms "include" and "comprise" used in the description and claims of the present invention indicate the presence of the described features, wholes, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or their collections.

[0043] It should also be understood that the terms used in this specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the specification of the present invention and the claims, the singular forms of "a", "an" and "the" are intended to include the plural forms unless the context clearly indicates otherwise. It should also be further understood that the term "and / or" used in the specification of the present invention and the claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes these combinations.

[0044] As used in this specification and claims, the term "if" may be interpreted as "when" or "upon" or "in response to determining" or "in response to detecting," depending on the context. Similarly, the phrase "if it is determined" or "if [described condition or event] is detected" may be interpreted as meaning "upon determination" or "in response to determining" or "upon detection of [described condition or event]" or "in response to detecting [described condition or event]," depending on the context.

[0045] The specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0046] Figure 1 An exemplary structural block diagram of an intelligent advertisement optimization delivery system 100 based on user behavior analysis according to an embodiment of the present invention is shown.

[0047] like Figure 1 As shown, in the intelligent advertisement optimization delivery system 100, the system includes:

[0048] Data collection and preprocessing module 101: used to collect user behavior data, and clean, analyze and process the user behavior data to obtain effective user behavior data; wherein the user behavior data includes browsing history, search history, and purchase behavior.

[0049] In one embodiment, the collecting of user behavior data, and cleaning, analyzing and processing the user behavior data to obtain effective user behavior data includes: data cleaning, feature extraction, data conversion and / or data desensitization based on user behavior data obtained from log files, databases and / or user interactions on the Internet.

[0050] In the example scenario, for example, an e-commerce platform, data sources may include the browsing history, search history, and purchase behavior of users on the platform. These data can be obtained from the log files, databases, and user interactions of the e-commerce platform website. The collected data may include structured data (such as purchase records), semi-structured data (such as web page elements in the browsing history), and unstructured data (such as user search queries).

[0051] In order to make it easier for those skilled in the art to understand, the steps involved in obtaining effective user behavior data are further described below. Among them, in the data cleaning, it can include: removing duplicate data, processing missing values, format unification and / or outlier detection and processing.

[0052] Removing duplicate data can be understood as deleting duplicate records in the database to ensure the uniqueness of each piece of data. Handling missing values ​​means that for missing data, you can choose to fill in (such as using the mean, median, or mode) or delete records with missing values. Then the data is formatted to ensure that all data follows the same format and standards, such as the unification of date and time formats. In addition, outlier detection and processing are also required. Statistical methods (such as the 3σ principle or box plot analysis) can be used to identify and process outliers, or outliers that exceed a certain range can be truncated or converted to numerical values.

[0053] Furthermore, the feature extraction may include: encoding conversion, word segmentation and part-of-speech filtering and / or constructing a feature vector. The encoding conversion may convert the original data into UTF-8 encoding to prevent encoding problems in subsequent processing. Word segmentation and part-of-speech filtering may be processed as word segmentation of the user's search records, removal of stop words, and retention of key information. Constructing a feature vector may be processed as converting text data into numerical features, such as using the TF-IDF method.

[0054] Furthermore, the data conversion may include: smoothing, aggregation, data generalization and / or data normalization. The smoothing process may smooth the noise in the data, such as using binning or regression methods. Aggregation may summarize the data, such as calculating the monthly or annual total. Data generalization may replace low-level data objects with more abstract concepts, such as replacing specific age values ​​with age attributes. Data normalization may scale attribute values ​​so that they fall into a specific interval, eliminating result deviations caused by different sizes of numerical data.

[0055] Furthermore, the data desensitization may include desensitizing the data containing the user's privacy, such as hiding the user's personal information and sensitive data to protect the user's privacy.

[0056] Furthermore, in the user portrait construction module 102 provided by the present invention: a user sentiment analysis unit and a user behavior pattern analysis unit are configured; they are used to construct a user portrait using the user's effective behavior data; the user portrait includes the user's attributes, interest preferences, and consumption habits. Among them, the user sentiment analysis unit is configured to use the user's effective behavior data to analyze the user's speech and emotions on social media to obtain the user's emotional response to the advertisement. The user behavior pattern analysis unit is configured to use the user's effective behavior data to identify the user type, and the classification dimensions of the user type include the consumption time dimension, the consumption frequency dimension, and the consumption amount dimension.

[0057] In one embodiment, the using of the user effective behavior data to analyze the user's speech and emotions on social media to obtain the user's emotional response to the advertisement includes:

[0058] First, a sentiment vocabulary is constructed to perform sentiment analysis on the text.

[0059] Secondly, use the bag-of-words model, TF-IDF model and / or word embedding model to extract features from text data.

[0060] Then, the Naive Bayes model is used to analyze the user's emotional tendency. The formula includes:

[0061]

[0062] Among them, P(C i |f 1 ,f 2 ,...,f n ) is a given feature f 1 ,f 2 ,...,f n Conditional emotion category C i The probability, P(Ci ) is the emotion category C i The prior probability, P(f i |C i ) is a given sentiment category C i Conditional feature f i The conditional probability of .

[0063] Or use the support vector machine formula to analyze the user's emotional tendency, the formula includes:

[0064] Among them, f(x) is the classification result of the input vector x, α i is the weight of the support vector, y i is the label of the support vector, K(x i ,x) is the kernel function and b is the bias term.

[0065] After the sentiment analysis model is completed, you can further use the validation set or test set to evaluate the model in order to optimize the model performance.

[0066] Further, in order to help those skilled in the art better understand the above solution, assume that there is a social media data set containing user comments on different advertisements. A naive Bayes model can be used to analyze the user's emotional tendency. The following is a specific embodiment pseudo code written in Python:

[0067] import numpy as np

[0068] from sklearn.naive_bayes import MultinomialNB

[0069] from sklearn.feature_extraction.text import TfidfVectorizer

[0070] from sklearn.model_selection import train_test_split

[0071] from sklearn.metrics import accuracy_score

[0072] #Assume that data is a dataset containing user comments and sentiment labels

[0073] comments = data['comment']

[0074] labels = data['sentiment']

[0075] #Feature extraction: Using TF-IDF model

[0076] vectorizer=TfidfVectorizer(stop_words='english')

[0077] X=vectorizer.fit_transform(comments)

[0078] y=labels

[0079] # Divide into training set and test set

[0080] X_train,X_test,y_train,y_test=train_test_split(X,y,test_size=0.2,random_state=42)

[0081] #Training Naive Bayes classifier

[0082] clf=MultinomialNB()

[0083] clf.fit(X_train,y_train)

[0084] # Make predictions on the test set

[0085] y_pred = clf.predict(X_test)

[0086] #Calculate accuracy

[0087] accuracy=accuracy_score(y_test,y_pred)

[0088] print("Accuracy:",accuracy)

[0089] Furthermore, in the user portrait construction module 102, the user effective behavior data is also used to identify the user type. The specific steps may include:

[0090] Step S4.1: extracting user behavior characteristics, including at least consumption time, consumption frequency, and consumption amount;

[0091] Step S4.2: Use K-means++ algorithm to select the initial cluster center;

[0092] Step S4.3: Use K-means algorithm to cluster user behavior patterns. The formula includes:

[0093]

[0094] Among them, J(C,μ) is the clustering quality function, which is used to evaluate the effect of clustering. That is, the sum of all clusters is added to get the clustering quality function J(C,μ). K is the number of clusters. Assume that the selected number of clusters is 3, that is, K = 3. C k Represents the set of all data points in the kth cluster. k is the center point of the kth cluster. x is a data point in the cluster; ||||x-μ k |||| 2 is the square of the Euclidean distance between the data point x and the cluster center.

[0095] Finally, the distribution of cluster centers and data points is updated iteratively until the value of J(C, μ) no longer changes significantly or reaches a preset number of iterations. The embodiment method provided by the present invention not only improves the accuracy of classification, but also enhances the explanatory power of the model.

[0096] Step S4.4: Construct a multi-layer perceptron (MLP) neural network for learning complex patterns of user behavior, wherein the multi-layer perceptron neural network comprises at least two hidden layers, wherein each hidden layer has 64 neurons. During the training process, the hidden layer and the softmax output layer of the ReLU activation function may be used, and the training may be performed using the Adam optimizer.

[0097] Use the activation function ReLU (Rectified Linear Unit), especially in the hidden layer. Its formula is:

[0098] f(x) = max(0, x), which means that if the input x is a positive number, the output of the ReLU function is x; if the input x is a negative number, the output is 0. The purpose of the ReLU function is to introduce nonlinearity so that the neural network can learn and simulate complex function mapping relationships. Compared with other activation functions (such as sigmoid or tanh), ReLU is smoother in the positive range, more efficient in calculation, and less likely to saturate during training.

[0099] For the Softmax function, it can be used in the output layer of a neural network, especially in multi-classification problems. Its formula can be:

[0100]

[0101] Where z is an input vector of length k, representing the raw scores of the output layer of the neural network; each element z i The score corresponding to class i.

[0102] K is the total number of categories, i.e. the number of different categories that the model needs to predict. i is the index of the category, indicating that the probability of the i-th category in the Softmax output is being calculated. is an exponential function that converts the original score z i Convert to a positive number to facilitate probability calculation. is the sum of the exponentials of all class scores. This sum acts as a normalization factor to ensure that the sum of all class probabilities equals 1.

[0103] Therefore, the calculation process of the Softmax function can be understood as follows:

[0104] First, an exponential function is applied to each element in the input vector z to convert the raw scores into positive numbers. This is because probabilities must be positive numbers, while the raw output of the neural network may contain negative values.

[0105] Then, the sum of all the exponentially scaled scores is calculated and used as a normalization factor to ensure that all elements in the output vector sum to 1.

[0106] Finally, each indexed score is divided by the sum, and the result is the probability of each class.

[0107] In this way, the Softmax function can convert the output of the neural network into a probability distribution, so that multiple categories can be predicted and the category with the highest probability can be selected as the final prediction result.

[0108] Furthermore, training is performed using the Adam optimizer, which calculates the first-order moment estimate (mean) and the second-order moment estimate (variance) of the gradient and combines these estimates to adjust the learning rate of each parameter.

[0109] The update rule of the Adam optimizer can be simplified as follows:

[0110] First, calculate the first-order moment estimate of the gradient (m t ) and the second-order moment estimate (v t ).

[0111] Second, bias correction is used to adjust these estimates.

[0112] Finally, the corrected estimates are used to update the parameters.

[0113] The technical effect of the above Adam optimizer is that it does not require manual adjustment of the learning rate and can automatically adapt the learning rate of each parameter, which enables it to perform well on many different problems.

[0114] Step S4.5: Fuzzify the output of the K-means algorithm and the prediction result of the multi-layer perceptron neural network and convert them into fuzzy sets; and

[0115] Using the fuzzy set, a fuzzy rule base is constructed to explain the relationship between the user behavior pattern and the prediction result of the multi-layer perceptron neural network;

[0116] Step S4.6: The K-means clustering results and the multi-layer perceptron neural network prediction results are weighted fused. The formula includes:

[0117] FinalBehavior=α·K-means output+(1-α)NeuralNetwork Prediction,

[0118] Among them, α is a weight parameter with a value range of 0 to 1. It is used to balance the importance of K-means clustering results and neural network prediction results. The specific value depends on the contribution of the two methods to the final classification results. If α is close to 1, it means that the K-means clustering results dominate the final decision; if α is close to 0, the prediction results of the neural network are more important.

[0119] K-means Output is the output of the K-means algorithm, which indicates the cluster label to which each user data point belongs. This output is classified based on the similarity of user behavior data.

[0120] Neural Network Prediction is the output of the multi-layer perceptron neural network model, which indicates the prediction result of the multi-layer perceptron neural network on user behavior. This output is a classification based on the complex patterns and relationships learned by the neural network.

[0121] Through the above-mentioned weighted fusion, the accuracy and robustness of classification can be improved.

[0122] The following is a further explanation using an example implementation. For example, there is a user behavior dataset of an e-commerce platform, which contains the user's purchase history, browsing history, and search history. The goal of this step is to classify users into different behavior pattern categories for more accurate marketing. Assume that through cross-validation, it is found that when α = 0.6, the classification effect of the model is the best. The pseudo code example implemented in Python is as follows:

[0123] from sklearn.cluster import KMeans

[0124] from sklearn.decomposition import PCA

[0125] from sklearn.neural_network import MLPClassifier

[0126] from sklearn.metrics import silhouette_score

[0127] #Assume that X is the preprocessed user behavior data

[0128] #Assume that y is the real user behavior category label (if any)

[0129] #PCA dimensionality reduction

[0130] pca = PCA (n_components = 0.95) # retain 95% of the variance

[0131] X_pca = pca.fit_transform(X)

[0132] #K-means clustering

[0133] kmeans=KMeans(n_clusters=3,init='k-means++',random_state=42)

[0134] kmeans_output=kmeans.fit_predict(X_pca)

[0135] #Neural network training

[0136] mlp=MLPClassifier(hidden_layer_sizes=(64,64), activation='relu', solver='adam', max_iter=1000)

[0137] mlp.fit(X_pca,y)#Assume that y is the real user behavior category label

[0138] #Neural network prediction

[0139] mlp_prediction=mlp.predict(X_pca)

[0140] #Model Fusion

[0141] alpha = 0.6

[0142] final_behavior=alpha*kmeans_output+(1-alpha)*mlp_prediction

[0143] #Evaluate clustering effect

[0144] silhouette_avg=silhouette_score(X_pca,kmeans_output)

[0145] print("Silhouette coefficient:",silhouette_avg)

[0146] Step S4.7: Use the fuzzy rule base to adjust the weights of different user behavior patterns to obtain clustering evaluation results, including using the silhouette coefficient index to evaluate the clustering effect. It should be noted that by using the fuzzy rule base to adjust the weights of different user behavior patterns, the importance of different patterns can be reflected.

[0147] Step S4.8: According to the clustering evaluation results, adjust the model parameters and fuzzy rules to optimize the type classification effect of the user behavior pattern.

[0148] Through the above embodiments, those skilled in the art can understand how to combine the K-means algorithm with other algorithms to classify user behavior patterns in a more detailed manner. This method not only improves the accuracy of classification, but also enhances the explanatory power of the model.

[0149] Furthermore, in the advertisement feature extraction module 103: it is used to perform multimodal feature extraction on the advertisement material to construct a semantic representation of the advertisement; wherein the multimodal feature extraction includes: extracting visual features using image recognition technology, and / or extracting text features using natural language processing technology.

[0150] In one embodiment, the use of image recognition technology to extract visual features includes: using the HOG algorithm to extract local shape features of the image; the steps include: converting the image into a grayscale image, calculating the gradient amplitude and direction of the image, constructing a histogram in a local unit according to the gradient direction, and normalizing the histogram to improve illumination invariance.

[0151] In a specific implementation scenario, because color information is not necessary for gradient calculation, and grayscale images can reduce the amount of calculation. The Sobel operator can be used to calculate the gradient amplitude and direction of the image. The Sobel operator estimates the gradient of the image by calculating the difference in the horizontal and vertical directions of the image. In normalizing the histogram to construct the histogram, the image can be divided into small cells, usually each cell is 8x8 pixels in size. In each cell, a histogram is constructed according to the gradient direction. Each bin (single bar or interval) of the histogram corresponds to a gradient direction range, which is used to count how many pixels have gradients in that direction, so that the algorithm can capture local shape features in the image and has good invariance to illumination changes. It can be divided into 9 bins from 0° to 180°, each bin covering a range of 20°. In order to improve illumination invariance, the histogram in each cell is normalized. This usually involves calculating the L2 norm of each cell histogram and normalizing it so that the total energy of the histogram remains consistent. Then, the histograms of all cells are concatenated to form the feature vector of each block. Usually, each block consists of several cells, for example, 2x2 cells form a block. Finally, the feature vectors of all blocks are concatenated to form the HOG feature vector of the entire image.

[0152] Through the above implementation steps, the HOG algorithm can effectively extract the local shape features of the image and construct a visual feature vector for describing the image. These feature vectors can be used for subsequent image recognition, classification and analysis tasks.

[0153] The key points and descriptors of the image are extracted using the SIFT algorithm; the steps include: detecting key points in spaces of different scales, assigning directions to the key points so that they are invariant to rotation, and constructing descriptors around the key points to capture local shapes and appearances.

[0154] Specifically, in scale space extremum detection, the SIFT algorithm first models the image in multi-scale space. By performing Gaussian blur operations at different scales, a series of scale spaces are obtained. In this scale space, the Difference of Gaussians (DoG) is used to detect local extreme points in the image, which are considered to be potential key points. Further, after the scale space extremum detection, the SIFT algorithm accurately locates the detected potential key points. By performing a second-order Taylor expansion on the neighborhood in the Gaussian pyramid, the position and scale of the key points are estimated, and key points with low contrast or unclear edge response are eliminated. Furthermore, in order to make the SIFT feature rotationally invariant, it is necessary to assign a dominant direction to each key point. Lowe proposed using the directional histogram of the local gradient of the image to determine the dominant direction of the key point. The gradient direction histogram is calculated in the local area around the key point, and the main peak in the histogram is selected as the dominant direction of the key point. After determining the position and direction of the key point, the SIFT algorithm will describe the local area around the key point. In one implementation scenario, a histogram descriptor based on the gradient around the key point, called the SIFT descriptor, can be used. The SIFT descriptor encodes the local gradient information of the image and normalizes the descriptor to enhance robustness. Finally, the position, scale, orientation and SIFT descriptor of the key point are combined to form a feature vector for each key point. These feature vectors are considered to be unique identifiers in the image and can be used for subsequent tasks such as image matching and target recognition.

[0155] The SIFT algorithm can effectively extract local shape features of an image and construct visual feature vectors for describing the image. These feature vectors can be used in the intelligent advertising optimization delivery system to improve the matching degree between advertising content and user interests, thereby improving advertising effectiveness and user experience.

[0156] Further, in the multimodal feature extraction of module 103, in one embodiment scenario, extracting text features using natural language processing technology includes:

[0157] The TF-IDF algorithm is used to extract keywords and their weights in the text. The formula includes:

[0158]

[0159]

[0160] X ij =TF(w j ,d i )×IDF(w j),

[0161] Among them, w j Indicates a specific term, d i Indicates specific text, n ij Indicates the term w j In text d i The number of times it appears in kj It is the entry w j The number of times it appears in all texts, max k n kj Indicates that in the text i The number of terms that appear most frequently in the j ,d i ) represents the term w j In text d i The frequency of occurrence in

[0162] IDF(w j ) is the term w j The inverse text frequency of N is the total number of texts in the corpus; n j Indicates that it contains the term w j The number of texts;

[0163] X ij Indicates the term w j In text d i TF-IDF value in;

[0164] Specifically, in one embodiment scenario, the TF-IDF algorithm is used to extract keywords and their weights in the text, including:

[0165] Text preprocessing: This includes removing stop words, punctuation, stemming, or word form restoration to reduce noise and standardize the text.

[0166] Build vocabulary: Create a vocabulary for all documents, containing all unique terms.

[0167] Calculate Term Frequency (TF): For each document, count the number of occurrences of each word in the vocabulary.

[0168] Calculate the Inverse Document Frequency (IDF): Calculate the IDF value of each word in the vocabulary to reflect the rarity of the word.

[0169] Calculate TF-IDF: Multiply TF and IDF to get the TF-IDF value of each word in each document.

[0170] Further, in one embodiment, the Word2Vec model is used to map words in the text to a vector space to capture the semantic relationship between words. The formula includes:

[0171]

[0172] Among them, v(w j ) is the term w j The vector representation of W is the weight matrix, e j It is the entry w j One-hot encoding of .

[0173] Specifically, in one embodiment scenario, using the Word2Vec model to map words in a text to a vector space includes:

[0174] Text preprocessing: Same as the above TF-IDF, necessary text cleaning and preprocessing are performed.

[0175] Build a corpus: Build the preprocessed text into a corpus to prepare for training the Word2Vec model.

[0176] Train the Word2Vec model: Use the Skip-gram or CBOW model to train the corpus and learn word vectors.

[0177] Word vector extraction: For each word in the corpus, obtain its corresponding vector representation.

[0178] By combining the above-mentioned solution of the present invention with TF-IDF and Word2Vec, the intelligent advertising optimization delivery system can extract rich features from the text, including the weight of keywords and the semantic relationship between words. These features can help the system understand the user's interests and advertising content more accurately, thereby improving the accuracy and effect of advertising delivery.

[0179] Furthermore, the intelligent advertising optimization delivery system based on user behavior analysis provided by the present invention also includes a conversion prediction model construction module 104: used to utilize the user effective behavior data and advertising features to construct a conversion prediction model by training a deep learning algorithm to obtain an advertising conversion evaluation effect; the advertising conversion evaluation effect includes the click-through rate, bid, and relevance factors of the advertisement.

[0180] The conversion prediction model building module of the present invention uses user effective behavior data and advertising features to predict the conversion effect of advertising, including click-through rate (CTR), conversion rate (CVR) and bid, etc. In the embodiment of the present invention, after preprocessing the user behavior data and advertising features by data cleaning, feature extraction and feature normalization, a fully connected network model is constructed by using a deep learning framework (such as TensorFlow or PyTorch). The model usually includes an input layer, multiple hidden layers and an output layer. Among them, in the input layer, preprocessed user behavior data and advertising feature vectors are received. In the hidden layer, one or more fully connected layers are used, and each neuron is connected to all neurons in the previous layer. The hidden layer may include an activation function (such as ReLU) to introduce nonlinearity. In the output layer, the number of neurons and activation functions of the output layer may be different depending on the task. For example, for a binary classification problem, the output layer may have only one neuron, and a sigmoid activation function may be used.

[0181] In one embodiment, the deep learning algorithm provided by the embodiment of the present invention includes:

[0182] A fully connected network is used to learn the relationship between user behavior data and advertising features. The formula includes:

[0183] P = FCN (F image ,F text ,U),

[0184] Among them, U is the user feature, P is the predicted conversion efficiency; F image is the extracted image feature, F text is the extracted text feature, and FCN is the fully connected network function.

[0185] In model training, you can use supervised learning methods to train a fully connected network model. This includes selecting a suitable loss function, such as cross entropy loss or mean square error loss, selecting an optimization algorithm, such as Adam or SGD, to update the model weights, and calculating the gradient of the loss function with respect to the model parameters, and updating the weights to minimize the loss function. Finally, evaluate the model performance on the validation set and adjust the model structure or hyperparameters as needed.

[0186] The above fully connected network can learn the complex relationship between user behavior data and advertising features, thereby improving the accuracy and effect of advertising delivery. Therefore, after the teachings of multiple embodiments of the present invention, those skilled in the art can understand that this model can be applied to various embodiment scenarios such as personalized recommendation and advertising click-through rate prediction.

[0187] Furthermore, the intelligent delivery decision module 105 of the present invention is used to construct an intelligent delivery decision model based on the user portrait, the semantic representation of the advertisement, and the advertisement conversion evaluation effect to generate an advertisement delivery strategy. This module includes: an intelligent sorting and matching unit: used to intelligently sort and match candidate advertisements based on the advertisement conversion evaluation effect using the LTP sorting model.

[0188] In one embodiment, the construction of the LTP sorting model includes:

[0189] First, we build an LTP ranking model to learn users’ preferences for advertisements and predict the ranking of advertisements. The formula includes:

[0190] S = f(U, A), where S is the ranking score of the ad, U is the user feature, A is the ad feature, and f is the function of the ranking model;

[0191] Secondly, the LTP sorting model is trained using a loss function, the formula includes:

[0192] L=∑ (i,j)∈D log(1+exp(-s i +s j ), where D is the set of text pairs in the training dataset, s i and j are the scores of text i and j respectively, and L represents the loss value.

[0193] In one embodiment, the LTP model can be implemented in the following ways: converting the sorting problem into a regression problem, predicting a score for each advertisement separately, and then sorting them according to the score; converting the sorting problem into a sorting pair comparison problem, such as using the RankNet model, and training the model by comparing the relative order of advertisement pairs; or considering the order of the entire advertisement list and directly optimizing the list-level evaluation indicators, such as LambdaRank and LambdaMART.

[0194] Using the above features and ranking model, you can train an intelligent delivery decision model. During model training, you can use machine learning frameworks such as LightGBM and set corresponding parameters, such as objective as lambdarank and metric as NDCG or MAP. The trained model is evaluated on the validation set to ensure that its prediction performance meets expectations. Evaluation indicators can include NDCG, MAP, etc., which can measure the effect of the model in actual applications.

[0195] Furthermore, the effect verification and optimization module 106 provided by the present invention is configured to include a verification unit and a delivery optimization unit; wherein the verification unit is configured to evaluate the delivery effects of different advertising strategies in real time through an A / B test online experiment mechanism to verify the estimated effect of the advertising delivery strategy. The delivery optimization unit is configured to optimize the advertising sorting and matching according to the delivery effect feedback of the verification unit to improve the intelligent delivery decision model.

[0196] In one embodiment, the verification unit evaluates the delivery effects of different advertising strategies in real time through A / B testing to verify the estimated effect of the advertising delivery strategy. The steps may include:

[0197] Formulate a hypothesis: Determine the test goal, such as increasing click-through rate or conversion rate, and set the original strategy A and the new strategy B.

[0198] Sample selection: Randomly select samples from the total users and divide them into two groups, one group uses the original strategy A and the other group uses the new strategy B.

[0199] Experiment Execution: Expose two groups of users to different advertising strategies and collect performance data.

[0200] Statistical analysis: Statistical methods (such as Z test) were used to analyze whether the difference between the two groups was significant.

[0201] Result evaluation: Based on the statistical analysis results, determine whether the new strategy B is better than the original strategy A.

[0202] Specific algorithm formulas may include:

[0203] Hypothesis Testing: H 0 (There is no difference between the original strategy and the new strategy) and H 1 (The new strategy is better than the original strategy)

[0204] Z-test formula: in, and is the mean of the two groups of samples, S 1 2 and S 2 2 is the sample variance, n 1 and n 2 is the sample size.

[0205] In one embodiment, the delivery optimization unit includes: using a gradient descent method to update model parameters to optimize the advertisement ranking, and the formula includes:

[0206] θ t+1 =θ t -η▽ θ J(θ t ), where θt represents the value of the model parameter at the tth iteration, θ t+1 represents the updated value of the model parameters at the t+1th iteration, η is the learning rate, which is used to control the step size of each step when updating the parameters, ▽ θ J(θ t ) is the gradient of the loss function J with respect to the parameter θ; J(θ t ) represents the loss function at the tth iteration.

[0207] Based on the embodiments of the technical solutions of the present invention described above, the specific implementation steps may include:

[0208] Data collection: Collect data such as advertising exposure, clicks, and conversions.

[0209] Model training: Use machine learning algorithms (such as deep learning) to predict users’ click-through rates and conversion rates.

[0210] Feature extraction: Extract useful features from user behavior data, such as the user's historical click-through rate, browsing time, etc.

[0211] Sorting logic: Calculate the score of each ad and sort them based on the predicted click-through rate and conversion rate, as well as the ad bid.

[0212] Parameter update: Use gradient descent to update model parameters to optimize ad ranking.

[0213] Effectiveness evaluation: Continuously monitor advertising effectiveness and adjust advertising strategies based on feedback.

[0214] In summary, the entire technical solution deeply integrates user portraits, ad semantic representation, and ad conversion evaluation effects, and uses advanced machine learning algorithms and big data analysis technologies to achieve intelligent ad placement decisions. This solution can accurately match user interests and ad content, optimize ad ranking, increase ad click-through rates and conversion rates, reduce invalid ad exposure, increase advertisers’ return on investment, and enhance user experience.

[0215] Based on the above, the present invention also discloses a computer-readable storage medium, on which is stored a computer program for an intelligent advertising optimization delivery system based on user behavior analysis. When the computer-readable program is executed by one or more processors, the module functions described in the previous multiple embodiments or implementation methods are implemented.

[0216] Although multiple embodiments of the present invention have been shown and described herein, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Those skilled in the art may conceive of many changes, modifications, and alternatives without departing from the thought and spirit of the present invention. It should be understood that in the process of practicing the present invention, various alternatives to the embodiments of the present invention described herein may be adopted. The appended claims are intended to define the scope of protection of the present invention, and therefore cover equivalents or alternatives within the scope of these claims.

Claims

1. Intelligent advertising optimization delivery system based on user behavior analysis, characterized by: The system comprises Data collection and preprocessing module: used to collect user behavior data, and clean, analyze and process the user behavior data to obtain effective user behavior data; wherein the user behavior data includes browsing history, search history, and purchase behavior; User portrait construction module: The configuration includes a user sentiment analysis unit and a user behavior pattern analysis unit; it is used to use the user's effective behavior data to construct a user portrait; the user portrait includes the user's attributes, interest preferences, and consumption habits; wherein, The user sentiment analysis unit is configured to analyze the user's speech and sentiment on social media using the user's effective behavior data to obtain the user's emotional response to the advertisement; The user behavior pattern analysis unit is configured to use the user effective behavior data to identify the user type, and the classification dimensions of the user type include consumption time dimension, consumption frequency dimension, and consumption amount dimension; Advertisement feature extraction module: used to extract multimodal features of advertisement materials to construct a semantic representation of advertisements; wherein the multimodal feature extraction includes: extracting visual features using image recognition technology, and / or extracting text features using natural language processing technology; Conversion prediction model building module: used to build a conversion prediction model by using the user's effective behavior data and advertising features, through training the deep learning algorithm, to obtain the advertising conversion evaluation effect; the advertising conversion evaluation effect includes the click-through rate, bid, and relevance factors of the advertisement; Intelligent delivery decision module: used to build an intelligent delivery decision model based on the user portrait, the semantic representation of the advertisement, and the advertisement conversion evaluation effect to generate an advertisement delivery strategy; this module includes: Intelligent sorting and matching unit: used to intelligently sort and match candidate advertisements using the LTP sorting model according to the advertisement conversion evaluation effect; Effect verification and optimization module: The configuration includes verification unit and delivery optimization unit; among them, The verification unit is configured to evaluate the delivery effects of different advertising strategies in real time through an A / B test online experiment mechanism to verify the estimated effect of the advertising delivery strategy; The delivery optimization unit is configured to optimize the advertisement sorting and matching according to the delivery effect feedback of the verification unit, so as to improve the intelligent delivery decision model.

2. The intelligent advertising optimization delivery system according to claim 1, characterized in that: The collecting of user behavior data, and cleaning, analyzing and processing the user behavior data to obtain effective user behavior data includes: Based on the user behavior data obtained from log files, databases and / or user interactions on the Internet, data cleaning, feature extraction, data conversion and / or data desensitization are performed; wherein, The data cleaning includes: removing duplicate data, processing missing values, format unification and / or outlier detection and processing; The feature extraction includes: encoding mode conversion, word segmentation and part-of-speech filtering and / or constructing a feature vector; The data conversion includes: smoothing processing, aggregation processing, data generalization and / or data normalization; The data desensitization includes: desensitizing the data containing user privacy.

3. The intelligent advertising optimization delivery system according to claim 1, characterized in that: The utilizing of the user effective behavior data to analyze the user's speech and emotions on social media to obtain the user's emotional response to the advertisement includes: Build a sentiment vocabulary to perform sentiment analysis on text; Use bag-of-words, TF-IDF, and / or word embedding models to extract features from text data; The Naive Bayes model is used to analyze the user's sentiment tendency. The formula includes: Among them, P(C i |f1,f2,...,f n ) for given features f1,f2,...,f n Conditional emotion category C i The probability, P(C i ) is the sentiment category C i The prior probability, P(f i |C i ) is a given sentiment category C i Conditional feature f i The conditional probability of .

4. The intelligent advertisement optimization delivery system according to claim 1, characterized in that: The step of using the user's effective behavior data to identify the user type includes: Step S4.1: extracting user behavior characteristics, including at least consumption time, consumption frequency, and consumption amount; Step S4.2: Use K-means++ algorithm to select the initial cluster center; Step S4.3: Use K-means algorithm to cluster user behavior patterns. The formula includes: Among them, J(C,μ) is the clustering quality function, which is used to evaluate the clustering effect; K is the number of clusters; C k represents the set of all data points in the kth cluster; μ k is the center point of the kth cluster; x is the data point in the cluster; ||||x-μ k |||| 2 is the square of the Euclidean distance between the data point x and the cluster center; Step S4.4: constructing a multi-layer perceptron neural network for learning complex patterns of user behavior, wherein the multi-layer perceptron neural network comprises at least two hidden layers; Step S4.5: Fuzzify the output of the K-means algorithm and the prediction result of the multi-layer perceptron neural network and convert them into fuzzy sets; and Using the fuzzy set, a fuzzy rule base is constructed to explain the relationship between the user behavior pattern and the prediction result of the multi-layer perceptron neural network; Step S4.6: The K-means clustering results and the multi-layer perceptron neural network prediction results are weighted fused. The formula includes: FinalBehavior=α·K-meansoutput+(1-α)NeuralNetworkPrediction, Among them, α is a weight parameter with a value range of 0 to 1; K-means Output is the output of the K-means algorithm, indicating the cluster label to which each user data point belongs; Neural Network Prediction is the output of the multilayer perceptron neural network model, indicating the prediction result of the multilayer perceptron neural network on user behavior; Step S4.7: using the fuzzy rule base to adjust the weights of different user behavior patterns to obtain clustering evaluation results; including using the silhouette coefficient indicator to evaluate the clustering effect; Step S4.8: According to the clustering evaluation results, adjust the model parameters and fuzzy rules to optimize the type classification effect of the user behavior pattern.

5. The intelligent advertisement optimization delivery system according to claim 1, characterized in that: The method of extracting visual features using image recognition technology includes: The HOG algorithm is used to extract local shape features of the image; the steps include: converting the image into a grayscale image, calculating the gradient amplitude and direction of the image, constructing a histogram in a local unit according to the gradient direction, and normalizing the histogram to improve illumination invariance; The key points and descriptors of the image are extracted using the SIFT algorithm; the steps include: detecting key points in spaces of different scales, assigning directions to the key points so that they are invariant to rotation, and constructing descriptors around the key points to capture local shapes and appearances.

6. The intelligent advertisement optimization delivery system according to claim 1, characterized in that: The extracting of text features using natural language processing technology includes: The TF-IDF algorithm is used to extract keywords and their weights in the text. The formula includes: X ij =TF(w j ,d i )×IDF(w j ) Among them, w j Indicates a specific term, d i Indicates specific text, n ij Indicates the term w j In text d i The number of times it appears in kj It is the entry w j The number of times it appears in all texts, max k n kj Indicates that in the text i The number of terms that appear most frequently in the j ,d i ) represents the term w j In text d i The frequency of occurrence in IDF(w j ) is the term w j The inverse text frequency of N is the total number of texts in the corpus; n j Indicates that it contains the term w j The number of texts; X ij Indicates the term w j In text d i TF-IDF value in; as well as, The Word2Vec model is used to map words in the text to vector space to capture the semantic relationship between words. The formula includes: Among them, v(w j ) is the term w j The vector representation of W is the weight matrix, e j It is the entry w j One-hot encoding of .

7. The intelligent advertisement optimization delivery system according to claim 1, characterized in that: The deep learning algorithm includes: A fully connected network is used to learn the relationship between user behavior data and advertising features. The formula includes: P=FCN(F image ,F text ,U), Among them, U is the user feature, P is the predicted conversion efficiency; F image is the extracted image feature, F text is the extracted text feature, and FCN is the fully connected network function.

8. The intelligent advertisement optimization delivery system according to claim 1, characterized in that: The construction of the LTP sorting model includes: Construct an LTP ranking model to learn users’ preferences for ads and predict the ranking of ads. The formula includes: S = f(U, A), where S is the ranking score of the ad, U is the user feature, A is the ad feature, and f is the function of the ranking model; The LTP sorting model is trained using a loss function, the formula includes: L=∑ (i,j)∈D log(1+exp(-s i +s j ), where D is the set of text pairs in the training dataset, s i and j are the scores of text i and j respectively, and L represents the loss value.

9. The intelligent advertisement optimization delivery system according to claim 1, characterized in that: The delivery optimization unit includes: Use the gradient descent method to update the model parameters to optimize the ad ranking. The formula includes: Among them, θ t represents the value of the model parameter at the tth iteration, θ t+1 represents the updated value of the model parameters at the t+1th iteration, η is the learning rate, which is used to control the step size of each step when updating the parameters. is the gradient of the loss function J with respect to the parameter θ; J(θ t ) represents the loss function at the tth iteration.

10. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by the processor, the intelligent advertising optimization delivery system as described in any one of claims 1-9 is implemented.

Images