E-commerce digital marketing advertisement pushing method and system based on distributed polyhedron matrix

Abstract

The invention discloses an e-commerce digital marketing advertisement pushing method and device based on a distributed polyhedron matrix. The method comprises the following steps: representing user and advertisement features as high-dimensional matrixes respectively; a matching process of user portraits and advertisement features is expressed as a geometric problem in a high-dimensional space, an advertisement pushing task is decomposed into a plurality of sub-tasks, and the sub-tasks are executed in parallel in a multi-core processor or a distributed system; performing matrix decomposition on the user portrait matrix and the advertisement feature matrix, extracting potential features of the user and the advertisement, calculating the matching degree between the user and the advertisement, and optimizing the calculation process; defining a target function, defining constraint conditions through a polyhedral model and representing the constraint conditions as linear inequalities, and solving an optimal advertisement pushing scheme through an iterative optimization algorithm in combination with matrix analysis and the polyhedral model; reinforcing learning is introduced to dynamically adjust a pushing strategy of the advertisement, and the pushing strategy is combined with a polyhedral model and matrix analysis to continuously learn an optimal pushing strategy; in the user portrait matrix, a social network polyhedral model is constructed in combination with multi-modal data based on the structure of the social network and the influence relationship between the users; data changes are captured in real time, user feedback is collected, a feature matrix is updated, unified construction is carried out in cooperation with multi-platform data, meanwhile, a self-adaptive algorithm parameter adjusting mechanism is designed, and related parameters are automatically adjusted according to the real-time changes of the data and the operation effect of an algorithm.

Description

Technical Field

[0001] The present invention belongs to the fields of computer science and technology and digital marketing, and is applied to the e-commerce industry. Specifically, it relates to a method for pushing e-commerce advertisements based on a distributed polyhedron matrix. Background Art

[0002] Currently, common advertisement recommendation algorithms in the e-commerce industry include content-based recommendation algorithms, large language model-based recommendation algorithms, collaborative filtering recommendation algorithms, and generative recommendation algorithms, etc. Among them, content-based recommendation algorithms recommend based on the matching degree between advertisement content features and users' historical preferences. They cannot mine users' implicit preferences and can only recommend based on the attributes of the advertisements themselves. The diversity of recommendations is limited, and long-term use may result in users only seeing similar types of advertisements. Moreover, it is difficult to accurately describe some complex and subjective advertisement features, such as creativity and emotion, which affects the accuracy of recommendations.

[0003] Large language model (LLM)-based recommendation algorithms can perform extensive pre-training on factual information, domain expertise, and common sense reasoning. Even without prior exposure to specific items or users, they can use their knowledge reserves to provide reasonable recommendations. However, their computational cost is high, the model parameter scale is huge, and a large amount of computational resources and time are required for reasoning and generation, resulting in a possible slow response speed of the recommendation system, especially when dealing with large-scale data and high-concurrency requests; and they are highly dependent on input data, and the recommendation effect largely depends on the quality and diversity of the input data. If the input data is biased, noisy, or incomplete, it may affect the accuracy and reliability of the recommendations; in addition, it is also difficult to process non-text data, and its processing ability for non-text data such as images and audio is relatively limited, and other technologies need to be combined for fusion processing.

[0004] Collaborative filtering recommendation algorithms recommend based on user behavior data, calculating the similarity between users or between items. However, user-based collaborative filtering is easily affected by data sparsity. When there is little user behavior data, it is difficult to find similar users, and there is also a cold start problem, and the recommendation effect is poor when new users or new advertisements are added; item-based collaborative filtering is easily affected by item popularity, and popular items are easily over-recommended.

[0005] Generative recommendation algorithms can generate brand-new and innovative recommendation results, breaking through the limitations of traditional recommendation algorithms based on historical data and similarity, and providing more personalized and unique recommendations for users. However, the quality of the generated results may be affected by various factors, such as the training data of the model, the training method, the generation strategy, etc., and there may be situations where the generated results are unreasonable, irrelevant, or of low quality; moreover, due to the diversity and uncertainty of the generated results, it is difficult to accurately evaluate them through metrics such as accuracy and recall like traditional recommendation algorithms, and dedicated evaluation methods and metrics need to be designed; at the same time, it has high requirements for data and computing resources. To generate high-quality recommendation results, a large amount of training data and powerful computing resources are required to support the training and inference of the model, otherwise overfitting or poor generation effects may occur. Summary of the Invention

[0006] The present invention aims to overcome the above-mentioned disadvantages of the prior art and provides an e-commerce digital marketing advertisement pushing method and system based on a distributed polyhedron matrix.

[0007] The first aspect of the present invention relates to an e-commerce digital marketing advertisement pushing method based on a distributed polyhedron matrix, including the following steps:

[0008] S1. Represent the characteristics of users and advertisements as a high-dimensional matrix respectively, where each row represents a user and each column represents a characteristic;

[0009] S2. Represent the matching process of the user portrait and advertisement characteristics as a geometric problem in a high-dimensional space through a polyhedron model, utilize the parallelization ability of the polyhedron model to decompose the advertisement pushing task into multiple subtasks, and execute them in parallel on a multi-core processor or a distributed system;

[0010] S3. Perform matrix decomposition on the user portrait matrix and the advertisement characteristic matrix to extract the latent characteristics of the users and advertisements and calculate the matching degree between the users and advertisements, and optimize the calculation process by using matrix analysis theory and deep learning techniques;

[0011] S4. Define an objective function and represent it as matrix operations, define constraint conditions through the polyhedron model and represent them as linear inequalities, and combine matrix analysis and the polyhedron model to solve the optimal advertisement pushing scheme through an iterative optimization algorithm;

[0012] S5. Introduce reinforcement learning to dynamically adjust the pushing strategy of the advertisement, combine it with the polyhedron model and matrix analysis, and continuously learn the optimal pushing strategy;

[0013] S6. Integrate the user's social network information into the user portrait matrix, and based on the structure of the social network and the influence relationship between users, combine multi-modal data to construct a social network polyhedron model;

[0014] S7. Real-time capture data changes and collect the user feedback for updating the feature matrix, collaboratively construct the unified data across multiple platforms, and meanwhile design an adaptive algorithm parameter adjustment mechanism to automatically adjust relevant parameters according to the real-time data changes and the running effect of the algorithm.

[0015] Preferably, the matrix representation of the user portrait and the advertisement features in step S1 includes:

[0016] Collect multi-dimensional feature data such as the user's interests, behaviors, and geographical locations, convert them into numerical forms to form a high-dimensional matrix, where each row corresponds to a user and each column represents a feature;

[0017] Quantify the features such as the category, target audience, and delivery time of the advertisement to form a matrix, where each row represents an advertisement and each column corresponds to a feature.

[0018] Preferably, the polyhedron model in step S2 is used for task scheduling, including:

[0019] Convert the matching conditions of the user portrait and the advertisement features into a set of linear inequalities, which define a polyhedron. The matching process of the user portrait and the advertisement features is equivalent to finding a suitable point in this polyhedron;

[0020] Utilize the characteristics of the polyhedron model to decompose the advertisement pushing task into multiple subtasks, allocate users in different regions or advertisements of different categories to different processors or nodes, and execute them in parallel in a multi-core processor or a distributed system.

[0021] Preferably, the matrix analysis in step S3 is used for personalized recommendation, including:

[0022] Perform matrix decomposition on the user portrait matrix and the advertisement feature matrix, including methods such as singular value decomposition (SVD) and non-negative matrix factorization (NMF), extract the latent features of the users and the advertisements, and calculate the matching degree between the users and the advertisements;

[0023] When extracting the features of the users and the advertisements, introduce a deep learning model to automatically learn more complex feature representations, and fuse the features extracted by the deep learning with the features in the original high-dimensional matrix;

[0024] When calculating the matching degree between the users and the advertisements, use a multi-layer perceptron MLP recommendation model based on deep learning and train the MLP on a large amount of user-advertisement interaction data.

[0025] Preferably, the design of the digital marketing algorithm in step S4 includes:

[0026] Determine the objective function of the advertisement push, including maximizing the user click-through rate (CTR) or return on investment (ROI), etc., and by constructing an objective function, use the user portrait matrix, advertisement feature matrix, and their matching degree matrix to calculate the expected CTR or ROI, and represent this objective in the form of matrix operations;

[0027] Define the constraint conditions of the advertisement push with the help of the polyhedron model, and represent these constraint conditions as linear inequalities;

[0028] Combine the matrix analysis and polyhedron model, and use the iterative optimization algorithm to solve the optimal advertisement push scheme, and continuously iteratively adjust the parameters in the user portrait matrix and advertisement feature matrix.

[0029] Preferably, the introduction of the reinforcement learning dynamic adjustment strategy in step S5 includes:

[0030] Regard the advertisement push process as a Markov decision process (MDP), where the state is the characteristics of the current user, advertisement inventory situation, and historical push effect, etc., the action is to select the advertisement to be pushed, and the reward is determined according to the user's feedback;

[0031] In the iterative optimization algorithm, use reinforcement learning to adjust the step size, direction of the iteration, or select different optimization paths.

[0032] Preferably, the integration of social network factors and multi-modal data processing in step S6 includes:

[0033] Analyze the user's social relationships, interest preferences of social groups, and social interaction behaviors, etc., and convert this information into quantifiable features and add them to the matrix;

[0034] Based on the structure of the social network and the influence relationship between users, consider the matching between the user's own characteristics and the advertisement, as well as the user's position in the social network and the potential interest of their friends or followers in the advertisement, and construct a social network polyhedron model;

[0035] In addition to the existing user and advertisement feature data, introduce multi-modal data such as pictures and videos. For the picture and video content in the advertisement, use image recognition and video analysis technologies to extract key visual features, and integrate these features with text features into the advertisement feature matrix. For users, also extract relevant features and incorporate them into the user portrait matrix;

[0036] During the matrix factorization and matching degree calculation processes, a multi-modal deep neural network (MM-DNN) is used to input the user and advertisement feature data and the different modal data into a multi-layer perceptron (MLP) network simultaneously for non-linear transformation of the network and parameter learning.

[0037] Preferably, the real-time data update, collaboration, and adaptive mechanism in step S7 includes:

[0038] During the advertisement pushing process, feedback data of the user on the advertisement is collected, including behaviors such as clicking, jumping, and purchasing. According to this feedback data, the user portrait and the advertisement feature matrix are adjusted and updated.

[0039] A real-time data update system is established. When the user has new purchasing behaviors or browsing records, these new information are incorporated into the matrix.

[0040] Features of the user or advertisement on each platform and device are extracted. Using the principal component analysis (PCA) dimensionality reduction algorithm, the user features collected from different platforms are mapped into the same low-dimensional space. The advertisement pushing decision is made based on these aligned features.

[0041] An online learning algorithm is used to track data changes in real time, process the new data in the form of a data stream, and dynamically update parameters such as the linear inequality parameters in the multihedron model, the parameters of matrix factorization, and the parameters of the iterative optimization algorithm.

[0042] The second aspect of the present invention relates to an e-commerce digital marketing advertisement pushing device based on a distributed multihedron matrix, including a memory and one or more processors. When the one or more processors execute the executable code stored in the memory, it is used to implement the e-commerce digital marketing advertisement pushing method and system of the present invention based on a distributed multihedron matrix.

[0043] The third aspect of the present invention relates to a computer-readable storage medium, on which a program is stored. When the program is executed by a processor, it implements the e-commerce digital marketing advertisement pushing method and system of the present invention based on a distributed multihedron matrix.

[0044] The advantages of the present invention are as follows: Traditional advertising recommendation algorithms more or less face pain points such as poor accuracy, low computational efficiency, weak adaptability, and poor comprehensiveness. In response to this, the present invention collects user interest, behavior and other data from multiple dimensions, constructs a high-dimensional user portrait matrix, and quantifies advertising features to form a matrix. During the matching process, matrix factorization is used to extract latent features, combined with a deep learning model to automatically learn complex features, and a multi-layer perceptron is used to calculate the matching degree. Social network factors are also considered to comprehensively improve the matching accuracy between advertisements and users and achieve personalized recommendation; the advertising push task is decomposed into multiple subtasks using a polyhedron model and assigned to different processors or nodes for parallel execution, giving full play to the advantages of multi-core processors or distributed systems, significantly improving the task processing efficiency, being able to process a large amount of user and advertising data in a short time, and meeting real-time requirements; reinforcement learning is introduced, regarding the advertising push as a Markov decision-making process, determining rewards based on user feedback, dynamically adjusting the iteration step size, direction or optimization path, enabling the algorithm to continuously optimize according to the actual situation, adapt to different application scenarios and changes in user behavior. At the same time, the real-time data update, collaboration and adaptive mechanism can adjust the user portrait and advertising feature matrix in a timely manner according to the user's real-time feedback and new behaviors, ensuring that the recommendation always fits the user's latest needs; social network factors are integrated, analyzing social relationships, group interests and interaction behaviors, constructing a social network polyhedron model, not only considering the matching between the user himself and the advertisement, but also taking into account the potential interests of friends or followers in the social network. In addition, multi-modal data processing is introduced, integrating visual features such as pictures and videos with text features, and using a multi-modal deep neural network for feature learning and matching degree calculation, enriching the recommendation basis from multiple dimensions and improving the recommendation effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 is a system overview diagram of the method of the present invention.

[0046] Figure 2 is a flowchart of the method of the present invention.

[0047] Figure 3 is a schematic diagram of the device of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0048] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0049] Embodiment 1

[0050] Refer to Figure 1 and Figure 2 , this embodiment relates to an e-commerce digital marketing advertising push method based on a distributed polyhedron matrix, including the following steps:

[0051] S1. Represent the characteristics of users and advertisements as a high-dimensional matrix respectively, where each row represents a user and each column represents a feature;

[0052] S2. Represent the matching process of the user profile and advertisement features as a geometric problem in a high-dimensional space through a polyhedron model. Utilize the parallelization ability of the polyhedron model to decompose the advertisement pushing task into multiple subtasks and execute them in parallel on a multi-core processor or a distributed system;

[0053] S3. Perform matrix factorization on the user profile matrix and the advertisement feature matrix to extract the latent features of the user and the advertisement and calculate the matching degree between the user and the advertisement. Optimize the calculation process by using matrix analysis theory and deep learning techniques;

[0054] S4. Define an objective function and represent it as matrix operations. Define the constraint conditions through the polyhedron model and represent them as linear inequalities. Combine matrix analysis and the polyhedron model to solve the optimal advertisement pushing scheme through an iterative optimization algorithm;

[0055] S5. Introduce reinforcement learning to dynamically adjust the pushing strategy of the advertisement, combine it with the polyhedron model and matrix analysis, and continuously learn the optimal pushing strategy;

[0056] S6. Integrate the social network information of the user into the user profile matrix. Based on the structure of the social network and the influence relationship between users, combine multi-modal data to construct a social network polyhedron model;

[0057] S7. Capture data changes in real time, collect the user feedback, and update the feature matrix. Collaborate with multi-platform data for unified construction. At the same time, design an adaptive algorithm parameter adjustment mechanism to automatically adjust relevant parameters according to the real-time changes of data and the running effect of the algorithm;

[0058] By analyzing the browsing history, search records, favorite content, etc. of the user on the e-commerce platform, determine the user's interest preferences for different product categories, brands, styles, etc. At the same time, collect the user's purchase behavior data, such as purchase frequency, purchase amount, purchase time interval, etc.; and the user's interaction behavior on the platform, such as the number of times of clicking on advertisements, staying time, whether to participate in promotional activities, etc.

[0059] Obtain the user's IP address or obtain their GPS location information through user authorization to determine the region, city level, business district, etc. where the user is located. Users in different regions may have different consumption habits and needs.

[0060] Convert the collected multi-dimensional feature data into numerical form. For categorical data, use the one-hot encoding method for quantization; for numerical data, directly use it or perform standardization processing to make it in the same order of magnitude for subsequent calculations.

[0061] Construct a high-dimensional user portrait matrix with each row corresponding to a user and each column representing a feature.

[0062] Clarify the category of the product promoted by the advertisement, and analyze the target population of the advertisement, including information such as age range, gender, occupation, etc. Record the time when the advertisement is planned to be launched, such as day or night, weekday or weekend, specific holidays, etc. The activity and consumption behavior of users vary at different time periods. Similarly, use appropriate quantification methods to convert these advertisement features into numerical forms, and construct an advertisement feature matrix with each row representing an advertisement and each column corresponding to a feature.

[0063] Determine the matching criteria between the user portrait and the advertisement features. Assume that setting the matching condition to be greater than a certain threshold is one of the matching conditions, then convert it into a linear inequality: user purchase amount - threshold ≥ 0.

[0064] For each matching condition, convert it into a similar linear inequality form. Multiple such linear inequalities together define a polyhedron. Combine the above linear inequalities to construct a polyhedron model. In this model, each point in the high-dimensional space represents a possible user-advertisement matching situation, and the points inside or on the boundary of the polyhedron represent the situations that meet the matching conditions.

[0065] According to the geographical location characteristics of users, divide users into different regions, and regard the matching task of users in each region with the advertisement as a subtask and assign it to different processors or nodes; according to the category of the advertisement, regard the matching tasks of different categories of advertisements with all users as subtasks respectively.

[0066] In a multi-core processor or distributed system, these subtasks can be executed in parallel. Utilize the parallelization ability of the polyhedron model to improve the processing efficiency of the advertisement push task, quickly find the user-advertisement combinations that meet the matching conditions, and achieve efficient advertisement push.

[0067] The principle of Singular Value Decomposition (SVD) is as follows: For a user portrait matrix U (with dimensions m×n, where m is the number of users and n is the number of features) and an advertisement feature matrix A (with dimensions p×n, where p is the number of advertisements), singular value decomposition can decompose the matrix into the product of three matrices, namely and where, U S and U A are orthogonal matrices, Σ S and Σ A are diagonal matrices, and the elements on the diagonal are singular values, arranged from largest to smallest. V S and V A are also orthogonal matrices.

[0068] Use the SVD function in the existing mathematical calculation library to perform singular value decomposition on the user portrait matrix U and the advertisement feature matrix A respectively. According to actual requirements, select the top k larger singular values and their corresponding singular vectors to retain, so as to achieve the purpose of dimensionality reduction while retaining the main feature information of the matrix.

[0069] Non-negative matrix factorization (NMF) attempts to decompose the non-negative matrices U and A into the product of two non-negative matrices, that is, U≈W H H U and A≈W A H A where the elements in W H 、H U 、W A and H A are all non-negative. This decomposition method can obtain a more practical latent feature representation because in many practical applications, features usually have non-negativity.

[0070] Initialize the non-negative matrices W H 、H U 、W A and H A using a random initialization method or some heuristic initialization methods (if for improving the stability and convergence speed of the algorithm), and use an iterative optimization algorithm to update W H 、H U 、W A and H A so that the matrix product after decomposition is as close as possible to the original matrix. In each iteration, calculate the update amount according to the objective function and update the matrix elements.

[0071] Targetedly select deep learning models such as convolutional neural network (CNN), recurrent neural network (RNN) or their variants (such as LSTM, GRU). If the user and advertisement features contain data such as images and texts, select CNN to process image data and use RNN or its variants to process text data.

[0072] Input the relevant data of users and advertisements into the selected deep learning model for training, so that the model can learn the feature representation in the data. After the model training is completed, extract the feature vectors from the middle layer or output layer of the model. These feature vectors represent the deep features of user and advertisement data. Integrate the features extracted by these deep learning with the features in the original high-dimensional matrix to form new feature vectors as a more comprehensive feature representation of users or advertisements.

[0073] Collect a large amount of user - ad interaction data, including feedback information such as whether the user clicks on the ad and whether the user purchases the product. Use 70% of the data as the training set to train the MLP model; 15% of the data as the validation set to adjust the model hyperparameters; and the remaining 15% of the data as the test set to evaluate the model's performance.

[0074] Build a multi - layer perceptron model using a deep learning framework. The dimension of the input layer of the model is the dimension of the feature vector after fusing user and ad features. Multiple hidden layers can be set, and each hidden layer contains a certain number of neurons. The dimension of the output layer is 1, which is used to predict the matching degree of the user to the ad.

[0075] Input the user - ad feature vectors in the training set into the MLP model. Continuously adjust the weights and biases of the model through the backpropagation algorithm and optimizer to minimize the error between the model's prediction result and the actual user feedback (such as whether the user clicks on the ad). During the training process, use the validation set to monitor the model's performance and avoid overfitting.

[0076] Use the trained MLP model. Input the new user - ad feature vectors into the model. The result output by the model is the matching degree between the user and the ad, which can be used to measure the potential interest degree of the user in the ad, thereby providing a basis for ad push.

[0077] Select target metrics, including maximizing the user click - through rate (CTR) and maximizing the return on investment (ROI). Among them, CTR is one of the important metrics to measure ad effectiveness and reflects the degree of user interest in the ad. The calculation formula is: CTR = number of ad clicks / number of ad impressions. ROI takes into account the relationship between ad placement cost and revenue and is a key metric for advertisers. The calculation formula is: ROI = (ad revenue - ad cost) / ad cost.

[0078] Use the user portrait matrix (denoted as U), the ad feature matrix (denoted as A), and their matching degree matrix (denoted as M) to construct the objective function. Assume that U is an m×n matrix, A is a p×n matrix, and M is an m×p matrix, where m is the number of users, p is the number of ads, and n is the number of features.

[0079] The objective function can be expressed as: where x ij is a decision variable representing whether to push ad j to user i, x ij ∈{0,1}. The meaning of this objective function is to maximize the total number of clicks on the pushed ads for all users by reasonably selecting which ads to push to which users.

[0080] The following defines the constraints for advertisement push with the help of a polyhedron model. The constraint based on user resource limitations is manifested as the limitation on the number of advertisements received by users: Assume that each user can receive at most k advertisements to ensure the user experience and avoid over-push. For each user i, it can be expressed as a linear inequality: The constraint based on advertisement resource limitations is manifested as the limitation on the number of advertisement placements: For some advertisements, there may be an upper limit on the number of placements due to reasons such as budget and inventory. Let the maximum number of placements of advertisement j be l j , then there is a constraint: The constraint based on matching conditions is transformed according to the matching conditions defined in the polyhedron model. For example, as mentioned in S2, the matching conditions between user portraits and advertisement features are transformed into linear inequalities, and these inequalities also constitute the constraints for advertisement push. For example, if a certain matching condition is the matching of the age range of the advertisement target audience and the user's age, let the age of user i be age i , the lower limit of the target age of advertisement j be min_age j , and the upper limit be max_age j , then there is a constraint: and (max_age j - age i ) ≥ 0. In the actual matrix operations and optimization process, these related conditions are incorporated into the constraint system.

[0081] The following selects an iterative optimization algorithm to solve the optimal advertisement push plan. One can choose the gradient descent algorithm and its variants, simulated annealing algorithm, genetic algorithm, etc. Then initialize the decision variable x ij . Adopt a random initialization method to assign an initial value of 0 or 1 to each x ij , or perform a more reasonable initialization according to some prior knowledge. For the initialization of other related parameters, if the gradient descent algorithm is used, the learning rate (step size) α needs to be initialized, which determines the amplitude of parameter update in each iteration. Setting the learning rate too large may cause the algorithm to fail to converge, while setting it too small will make the convergence speed too slow. Therefore, a relatively large initial value is set in advance and then adjusted according to the convergence situation during the iteration process.

[0082] For the objective function J(x) (x represents the vector composed of all decision variables x ij ), calculate its gradient with respect to x . Consider the combined influence of the objective function and constraint conditions during the calculation process, and obtain the gradient expression through matrix operations and derivative rules.

[0083] Update the decision variable x according to the selected iterative optimization algorithm and the corresponding update rules. After each update, check whether the decision variable satisfies all the constraint conditions. If not, make corresponding adjustments, such as projecting it into the feasible region.

[0084] Set the convergence conditions, such as the change in the objective function value is less than a certain threshold ε, or the number of iterations reaches a preset maximum value T. When the convergence conditions are met, stop the iteration. At this time, the value of the decision variable x ij is the optimal advertising push scheme, that is, it determines which advertisements to push to which users to maximize the objective function.

[0085] Next, introduce reinforcement learning to dynamically adjust the advertising push strategy. First, it is necessary to define the Markov decision process (MDP). Collect multi-dimensional features of the current user, including browsing history, purchase records, interest preferences and other data, and combine the advertising inventory situation, including the remaining delivery volume of different types of advertisements, delivery time limits and other information, as well as historical push effects, including indicators such as the click-through rate and conversion rate of users on the pushed advertisements in the past period of time. Integrate this information into a state vector. Select an advertisement from the advertising inventory to push to the current user as an action. Let A t represent the advertisement selected at time t, and its value range is the set of all deliverable advertisements. Determine the reward according to the user's feedback on the pushed advertisement. If the user clicks on the advertisement, give a positive reward; if the user purchases the product promoted by the advertisement, give a higher positive reward; if the user does not respond, give a negative reward. The reward function R(S t ,A t ) represents the reward obtained by performing the action A t in the state S t .

[0086] Select reinforcement learning algorithms such as Q-Learning and Deep Q-Network (DQN) to build a deep neural network. Its input is the state vector S t , and the output is the Q value of each possible action (i.e., each advertisement). Use the experience replay mechanism to store each state transition (S t ,A t ,R t ,S t+1 ) into the experience replay pool, and randomly sample a batch of samples from the experience replay pool for training the neural network to update the Q value estimate.

[0087] In the iterative optimization algorithm, according to the current state S t , select the action A t, specifically using the ε-greedy strategy, that is, randomly selecting actions with a probability of ε, and selecting the action with the largest Q value with a probability of (1 - ε). As the iteration progresses, gradually decrease the value of ε to make the algorithm gradually shift from exploration to exploitation. According to the reward and the next state S t+1 , update the parameters of the neural network according to the update rules of the reinforcement learning algorithm, and continuously learn the optimal push strategy.

[0088] Then integrate the user's social network information into the user portrait matrix to construct a social network polyhedron model. Obtain data such as the user's friend list, follow and being followed relationships on the social platform, analyze the topics discussed and content shared in the social groups where the user belongs, infer the interest preferences of the groups, and count the frequency and objects of the user's interaction behaviors such as likes, comments, and forwards on the social platform. Convert this social network information into quantifiable features.

[0089] Add the quantified social network features to the original user portrait matrix. Assume the original user portrait matrix is U with a dimension of m×n, and the number of newly added social network features is k. Then the dimension of the extended user portrait matrix becomes m×(n + k).

[0090] Consider the relationship between the social network and influence. Based on the adjacency matrix of the social network, analyze the influence propagation paths between users, and calculate the influence weights of each user in the social network through algorithms such as PageRank. Combining the user's own characteristics with the advertisement matching, when considering advertisement pushing, not only consider the matching degree between the user's own characteristics and the advertisement, but also consider the user's position in the social network and the potential interest of their friends or followers in the advertisement.

[0091] In addition to social network information, introduce multimodal data such as pictures and videos. For the picture and video content in the advertisement, use image recognition and video analysis technologies to extract key visual features; for the content such as pictures and videos generated by users, also extract corresponding features, and integrate these visual features with text features into the advertisement feature matrix and the user portrait matrix.

[0092] According to the above factors, construct a social network polyhedron model, convert various relationships and features in the social network into linear inequality constraints, combine them with the original user portrait and advertisement feature matching conditions, and define a new polyhedron in the high-dimensional space for more accurate advertisement pushing decisions.

[0093] During the advertisement pushing process, build a real-time data collection system to comprehensively collect the feedback data of users on the advertisement. By embedding monitoring codes in the advertisement display page and related links, track key behavior data such as whether the user clicks on the advertisement, the jump behavior after clicking, and whether the purchase is finally completed.

[0094] According to the collected feedback data, the user portrait and the advertisement feature matrix are adjusted in a timely manner. If a user clicks on an advertisement for a certain type of technology product, in the user portrait matrix, the eigenvalue weight of the user's interest in technology products is correspondingly increased; if the click-through rate and conversion rate of an advertisement are much lower than expected, in the advertisement feature matrix, relevant eigenvalue such as the attractiveness of the advertisement and the matching degree of the target audience are adjusted. An incremental update method is adopted. Based on the original matrix, the adjustment amount is calculated according to the feedback data, and the matrix elements are updated in real time.

[0095] When a user generates new purchase behaviors or browsing records, these new information are quickly connected to the system through the data interface. The newly obtained data are integrated with the original user portrait and advertisement feature matrix. Through data cleaning and preprocessing, the accuracy and consistency of the new data are ensured, and then the matrix is updated according to the established rules.

[0096] Extract the features of users or advertisements from different platforms and devices. Using the principal component analysis (PCA) dimensionality reduction algorithm, map the high-dimensional user features collected from different platforms to the same low-dimensional space. First, calculate the covariance matrix of the feature data, then solve the eigenvalues and eigenvectors of the covariance matrix, select the eigenvectors corresponding to the top k largest eigenvalues, and project the original features into the low-dimensional space spanned by these eigenvectors to achieve feature alignment. According to the aligned low-dimensional features, make advertisement push decisions, and determine the content and order of the pushed advertisements by analyzing the similarity and matching degree between the user features and advertisement features in the low-dimensional space.

[0097] Adopt the online learning version of stochastic gradient descent (SGD) to track the dynamic changes of data in real time. Whenever a new data sample arrives, the newly arrived data are processed in the form of a data stream, the gradient of the sample is immediately calculated, and the model parameters are updated according to the gradient.

[0098] Dynamically update the parameters of the linear inequalities in the polyhedron model, the parameters of matrix factorization, and the parameters of the iterative optimization algorithm, etc. When it is found that the data distribution has changed significantly, adjust the constraint conditions of the polyhedron model in a timely manner, recalculate the results of matrix factorization, and optimize the step size and convergence conditions of the iterative optimization algorithm, etc., to ensure that the digital marketing algorithm can always adapt to the real-time changes of data and maintain a good push effect.

[0099] Embodiment 2

[0100] Refer to Figure 3 , this embodiment relates to an e-commerce digital marketing advertisement pushing device based on a distributed polyhedron matrix, including a memory and one or more processors. Executable code is stored in the memory. When the one or more processors execute the executable code, it is used to implement the e-commerce digital marketing advertisement pushing method based on the distributed polyhedron matrix in Embodiment 1.

[0101] Example 3

[0102] This embodiment relates to a computer-readable storage medium, on which a program is stored. When the program is executed by a processor, it implements the e-commerce digital marketing advertisement pushing method based on a distributed polyhedron matrix in Embodiment 1.

[0103] The content described in the embodiments of this specification is only an enumeration of the implementation forms of the inventive concept. The protection scope of the present invention should not be regarded as limited to the specific forms stated in the embodiments. The protection scope of the present invention also covers equivalent technical means that can be conceived by those skilled in the art based on the inventive concept of the present invention.

Claims

1. An e-commerce digital marketing advertisement pushing method based on a distributed polyhedron matrix, characterized in that, Including the following steps: S1. Represent the characteristics of users and advertisements as high-dimensional matrices respectively, where each row represents a user and each column represents a characteristic; S2. Represent the matching process of the user profile and advertisement characteristics as a geometric problem in a high-dimensional space through a polyhedron model. Utilize the parallelization ability of the polyhedron model to decompose the advertisement pushing task into multiple subtasks and execute them in parallel on a multi-core processor or a distributed system; S3. Perform matrix decomposition on the user profile matrix and advertisement characteristic matrix to extract the latent characteristics of the users and advertisements and calculate the matching degree between the users and advertisements. Optimize the calculation process using matrix analysis theory and deep learning techniques; S4. Define an objective function and represent it as matrix operations. Define constraint conditions through the polyhedron model and represent them as linear inequalities. Combine matrix analysis and the polyhedron model to solve the optimal advertisement pushing scheme through an iterative optimization algorithm; S5. Introduce reinforcement learning to dynamically adjust the pushing strategy of the advertisement. Combine it with the polyhedron model and matrix analysis to continuously learn the optimal pushing strategy; S6. Integrate the social network information of the users into the user profile matrix. Based on the structure of the social network and the influence relationship between users, combine multi-modal data to construct a social network polyhedron model; S7. Capture data changes in real time and collect user feedback to update the characteristic matrix. Collaborate with multi-platform data for unified construction. At the same time, design an adaptive algorithm parameter adjustment mechanism to automatically adjust relevant parameters according to the real-time changes of data and the running effect of the algorithm.

2. The e-commerce digital marketing advertisement pushing method based on a distributed polyhedron matrix according to claim 1, wherein The matrix representation of the user profile and advertisement characteristics in step S1 includes: Collect multi-dimensional characteristic data such as the interests, behaviors, and geographical locations of users, convert them into numerical forms, and form a high-dimensional matrix. Each row corresponds to a user and each column represents a characteristic; Quantify the characteristics such as the category, target audience, and placement time of the advertisement to form a matrix. Each row represents an advertisement and each column corresponds to a characteristic.

3. The e-commerce digital marketing advertisement pushing method based on a distributed polyhedron matrix according to claim 1, wherein The polyhedron model in step S2 is used for task scheduling and includes: Convert the matching conditions of the user profile and advertisement characteristics into a set of linear inequalities, which define a polyhedron. The matching process of the user profile and advertisement characteristics is equivalent to finding a suitable point in this polyhedron; Utilize the characteristics of the polyhedron model to decompose the advertisement pushing task into multiple subtasks, allocate users in different regions or advertisements of different categories to different processors or nodes, and execute them in parallel on a multi-core processor or a distributed system.

4. The e-commerce digital marketing advertisement pushing method based on a distributed polyhedron matrix according to claim 1, characterized in that The matrix analysis in step S3 is used for personalized recommendation and includes: Perform matrix decomposition on the user profile matrix and advertisement characteristic matrix, including the singular value decomposition SVD and non-negative matrix factorization NMF methods, extract the latent characteristics of the users and advertisements, and calculate the matching degree between the users and advertisements; When extracting the characteristics of users and advertisements, introduce a deep learning model to automatically learn more complex characteristic representations and fuse the characteristics extracted by the deep learning with the characteristics in the original high-dimensional matrix; When calculating the matching degree between users and advertisements, a multi-layer perceptron (MLP) recommendation model based on deep learning is used to train the MLP on a large amount of user-advertisement interaction data.

5. The e-commerce digital marketing advertisement pushing method based on a distributed polyhedron matrix according to claim 1, characterized in that, The design of the digital marketing algorithm described in step S4 includes: Determine the objective function of the advertisement push, including maximizing the user click-through rate (CTR) or return on investment (ROI), etc., and by constructing an objective function, use the user portrait matrix, advertisement feature matrix, and the matching degree matrix between them to calculate the expected CTR or ROI, and represent this objective in the form of matrix operations; Define the constraints of the advertisement push with the help of the polyhedron model, and represent these constraints as linear inequalities; Combining the matrix analysis and the polyhedron model, use an iterative optimization algorithm to solve the optimal advertisement push scheme, and continuously iteratively adjust the parameters in the user portrait matrix and advertisement feature matrix.

6. The e-commerce digital marketing advertisement pushing method based on a distributed polyhedron matrix according to claim 1, wherein The introduction of the reinforcement learning dynamic adjustment strategy described in step S5 includes: Regard the advertisement push process as a Markov decision process (MDP), where the state is the characteristics of the current user, the advertisement inventory situation, and the historical push effect, etc., the action is to select the advertisement to be pushed, and the reward is determined according to the user's feedback; In the iterative optimization algorithm, use reinforcement learning to adjust the step size, direction of the iteration, or select different optimization paths.

7. The e-commerce digital marketing advertisement pushing method based on a distributed polyhedron matrix according to claim 1, characterized in that The integration of social network factors and multi-modal data processing described in step S6 includes: Analyze the user's social relationships, the interest preferences of social groups, and social interaction behaviors, etc., and convert this information into quantifiable features and add them to the matrix; Based on the structure of the social network and the influence relationship between users, consider the matching between the user's own characteristics and the advertisement, as well as the user's position in the social network and the potential interest of their friends or followers in the advertisement, and construct a social network polyhedron model; In addition to the existing user and advertisement feature data, introduce multi-modal data such as pictures and videos. For the picture and video content in the advertisement, use image recognition and video analysis technologies to extract key visual features, and integrate these features with the text features into the advertisement feature matrix. For users, also extract relevant features and incorporate them into the user portrait matrix; In the matrix decomposition and matching degree calculation process, use a multi-modal deep neural network (MM-DNN) to input the user and advertisement feature data and the different modal data into a multi-layer perceptron (MLP) network at the same time for non-linear transformation and parameter learning of the network.

8. The e-commerce digital marketing advertisement pushing method and system based on a distributed polyhedron matrix according to claim 1, wherein The real-time data update, collaboration, and adaptive mechanism described in step S7 includes: During the advertisement push process, collect the feedback data of users on the advertisement, including behaviors such as clicks, jumps, and purchases. According to this feedback data, adjust and update the user portrait and advertisement feature matrix; Establish a real-time data update system. When the user has new purchase behaviors or browsing records, integrate this new information into the matrix; Extract the features of the user or advertisement on each platform and device, and use the principal component analysis (PCA) dimensionality reduction algorithm to map the user features collected from different platforms into the same low-dimensional space, and make the advertisement push decision according to these aligned features; Use the online learning algorithm to track the data changes in real time, process the new data in the form of data stream, and dynamically update the parameters of the linear inequalities in the polyhedron model, the parameters of matrix factorization, and the parameters of the iterative optimization algorithm, etc.

9. An e-commerce digital marketing advertisement pushing device based on a distributed polyhedron matrix, characterized in that It includes a memory and one or more processors. An executable code is stored in the memory. When the one or more processors execute the executable code, it is used to implement the e-commerce digital marketing advertisement push method based on the distributed polyhedron matrix according to any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, A program is stored thereon. When the program is executed by a processor, it implements the e-commerce digital marketing advertisement push method based on the distributed polyhedron matrix according to any one of claims 1-8.

Images