Advertisement putting engine generation method

Abstract

The invention relates to the technical field of advertisement putting engine generation methods, in particular to an advertisement putting engine generation method which specifically comprises the following steps: step 1, firstly, collecting multi-dimensional data of users through multiple channels, including but not limited to basic information of the users, and setting a collected user data set as U = {u1, u2,..., un}, where n represents the number of the users, and n represents the number of the users; ui represents all data of the ith user; and 2, carrying out user dynamic interest modeling by adopting a recurrent neural network model based on an attention mechanism. According to the invention, the complex potential relationship between the user and the advertisement can be fully mined.

Description

Technical Field

[0001] The present invention relates to the technical field of advertisement delivery engine generation methods, and in particular to an advertisement delivery engine generation method. Background Art

[0002] In the era of digital marketing, advertising is a crucial tool for companies to promote their products and services, and its effectiveness directly impacts their market competitiveness and economic returns. Traditional ad delivery engines often make decisions based on simple user profiles and historical clickthrough data. For example, early ad delivery systems primarily used a rule-based approach to match ads based on basic user information such as age, gender, and region. Specifically, ads were delivered to users whose attributes matched pre-defined rules. This approach is limited by its rigid rules, making it difficult to adapt to complex and changing user behavior and market conditions.

[0003] With technological advancements, machine learning-based ad delivery engines have emerged, such as the click-through rate prediction model based on logistic regression. While these models have improved ad delivery accuracy to a certain extent, they still face some challenges. For one thing, they fail to capture the dynamics of user behavior and struggle to reflect rapid shifts in user interests in real time. Furthermore, these models perform poorly with high-dimensional, sparse data, failing to fully tap into the complex underlying relationships between users and ads. Summary of the Invention

[0004] In order to solve the technical problem that the existing technology cannot fully explore the complex potential relationship between users and advertisements, the present invention provides a method for generating an advertisement delivery engine.

[0005] The technical solution adopted by the present invention is: a method for generating an advertisement delivery engine, which specifically includes the following steps:

[0006] Step 1: First, collect multi-dimensional data of users through various channels, including but not limited to basic information of users. Let the collected user data set be U = {u1,u2,…,u n}, where n represents the number of users, u i Represents all data of the i-th user;

[0007] Step 2: Use a recurrent neural network model based on the attention mechanism to model user dynamic interests;

[0008] Step 3: Extract features from the advertising data, including content features, visual features, and delivery features;

[0009] Step 4: Calculate the matching degree between users and ads by combining cosine similarity and deep matching network;

[0010] Step 5: Use ensemble learning methods to predict advertising effectiveness;

[0011] Step 6: Based on the predicted results of advertising effectiveness, a reinforcement learning method is used to dynamically allocate advertiser budgets;

[0012] Step 7: Based on the user and ad matching, ad delivery effect prediction and budget allocation results;

[0013] Step 8: Execute the advertising operation on the advertising platform according to the generated advertising strategy;

[0014] Step 9: Use multi-dimensional evaluation indicators to evaluate the effectiveness of advertising;

[0015] Step 10: Optimize the advertising delivery strategy based on the advertising effectiveness evaluation results.

[0016] In one embodiment, in step 1, the collected data is preprocessed, including data cleaning, missing value processing and normalization operations;

[0017] For missing values, multiple imputation method is used as follows:

[0018] Assume that user feature x j If there are missing values, first build a prediction model based on other complete features to predict the missing values, then generate m filled data sets, conduct subsequent analysis for each filled data set, and finally combine the m results to get the final conclusion;

[0019] Data normalization uses the minimum-maximum normalization method, and the formula is:

[0020]

[0021] where x ij represents the jth feature value of the i-th user, x′ ij Indicates the normalized eigenvalue, min(x j ) and max(x j ) represent the minimum and maximum values ​​of the j-th feature respectively.

[0022] In one embodiment, in step 2, the specific method of using a recurrent neural network model based on an attention mechanism to model user dynamic interests is as follows:

[0023] First, modeling is performed based on the long short-term memory network. The specific steps are as follows:

[0024] Assume that the user's behavior sequence is S = {s1, s2, ..., s t}, where s t represents the user's behavior at time t;

[0025] The input gate formula of the long short-term memory network is: t =σ(W xi s t +W hi h t-1 +W ci c t-1 +b i ); the forget gate formula is: f t =σ(W xf s t +W hf h t-1 +W cf c t-1 +b f ); the output gate formula is: t =σ(W xo s t +W ho h t-1 +W co c t-1 +b o ); the cell state update formula is: c t =f t ⊙c t-1 +i t ⊙tanh(W xc s t +W hc h t-1 +b c ), the hidden state update formula is: h t =o t ⊙tanh(c t ); where σ is the Sigmoid function, W xi ,W hi ,W ci etc. is the weight matrix, b i ,b f ,b o ,b c is the bias vector, ⊙ represents the element-wise product;

[0026] Introducing the attention mechanism, the formula for calculating the attention weight is: Where score(h i ,h j ) is used to calculate the hidden state h i and h j function of the correlation between them.

[0027] In one embodiment, in step 3, for the text content features of the advertisement, a word vector model is used to map each word into a low-dimensional vector. Let the word set in the advertisement text be W = {w1, w2, ..., wm}, get the word vector e(w i ), and then the average pooling method is used to obtain the vector representation of the advertising text

[0028] For the visual features of advertisements, a convolutional neural network is used to extract them. The method is as follows:

[0029] Let the advertisement image be I, and after the multi-layer convolution and pooling operation of the CNN network, the feature vector a is obtained. visual ;

[0030] Finally, the various feature vectors of the advertisement are concatenated to obtain the comprehensive feature representation vector of the advertisement:

[0031] a=[a text ;a visual ;a 投放特征 ].

[0032] In one embodiment, in step 4, based on the user dynamic interest representation vector v and the advertisement comprehensive feature representation vector a, a method combining cosine similarity and deep matching network is used to calculate the user-advertisement matching degree as follows:

[0033] The cosine similarity formula is:

[0034] Where v·a represents the vector dot product, ||v|| and ||a|| represent the norms of vectors v and a respectively;

[0035] The deep matching network adopts a multi-layer perceptron structure. The input is the concatenated vector [v; a], which passes through multiple hidden layers h1, h2, ..., h l The output matching score is y deep ;

[0036] The hidden layer calculation formula is:

[0037] h k =σ(W k h k-1 +b k )(k=1,2,…,l);

[0038] The output layer formula is:

[0039] Final user-ad matching score score = α·sim cos (v,a)+(1-α)·y deep ;

[0040] Among them, α is the weight coefficient.

[0041] In one embodiment, in step 5, the specific method for using the ensemble learning method to predict the effect of advertising delivery is as follows:

[0042] Combine multiple different prediction models, including gradient boosted decision trees, support vector machines, and neural networks;

[0043] Suppose the predicted click rate of user i seeing ad j by the kth prediction model is The final click-through rate prediction value is obtained by weighted average:

[0044] Where K is the number of models, w k is the weight of the k-th model, and w k Determined by cross-validation optimization;

[0045] Suppose the behavior sequence after the user clicks the ad is B={b1,b2,…,b s}, using the sequence-to-sequence (Seq2Seq) model to predict the conversion rate r ij .

[0046] In one embodiment, in step 6, based on the prediction results of the advertising effect, a reinforcement learning method is used to dynamically allocate the advertiser's budget. The specific method is as follows:

[0047] The advertising delivery process is modeled as a Markov decision process. State s represents the current advertising delivery status, including the remaining budget, the effectiveness of the delivered ads, and the characteristics of the user group. Action a represents the budget allocation strategy, and reward r represents the benefit obtained after executing the action.

[0048] The deep Q-network algorithm is used to learn the optimal budget allocation strategy. The Q-value function is expressed as Q(s, a; θ), where θ is a network parameter. The Q-value is updated according to the Bellman equation by continuously interacting with the environment:

[0049]

[0050] Here, α is the learning rate and γ is the discount factor.

[0051] In one embodiment, in step 7, based on the user and advertisement matching degree, advertisement delivery effect prediction, and budget allocation result, an advertisement delivery strategy is generated as follows:

[0052] First, the advertisements are sorted according to the matching degree and the predicted effect. Let the sorted advertisement list be A={a1,a2,…,a n};

[0053] Then, based on the budget allocation results, determine the budget, delivery time, and platform parameters for each ad.

[0054] For each user, select the ads with high matching degree and budget allowed from the sorted ad list for delivery;

[0055] Assume that user i's budget is budget i , the estimated cost of advertisement j is cost j , when budget i ≥cost j When the matching degree of ad j is above the threshold, ad j is delivered to user i and the budget of user i is updated. i =budget i -cost j .

[0056] In one embodiment, in step 8, the specific method for performing the advertisement delivery operation on the advertisement delivery platform according to the generated advertisement delivery strategy is as follows:

[0057] During the delivery process, user feedback data is collected in real time, including ad impressions, clicks, conversion rates, and user dwell time;

[0058] Let the collected impressions of ad j be impressions j , clicks j , the conversion rate is r ates j , the user stay time is time s pen.

[0059] In one embodiment, in step 9: the specific method of using multi-dimensional evaluation indicators to evaluate the advertising effect is as follows:

[0060] These include click-through rate (CTR), conversion rate (CVR), and return on investment (ROI). j , user engagement;

[0061] The click-through rate formula is:

[0062] The conversion rate formula is:

[0063] The ROI formula is:

[0064] At the same time, the entropy weight method is introduced to determine the weight of each evaluation indicator;

[0065] Let the jth index value of the i-th sample be x ij ; After normalization: The entropy value of the jth indicator Weight

[0066] Finally, the comprehensive evaluation score of advertisement j is obtained by combining the scores and weights of each indicator. Evaluate the effectiveness of advertising.

[0067] The beneficial effects of the present invention are as follows: compared with the existing technology, the advertising delivery engine generation method proposed in the present invention realizes accurate modeling of users' dynamic interests and effective extraction of advertising features through multi-dimensional data collection and preprocessing, combined with advanced deep learning and machine learning algorithms, and can capture the complex potential relationship between users and advertisements in real time and accurately, significantly improving the user-advertising matching degree. Advertisement delivery effect prediction and budget dynamic allocation are carried out by using methods such as ensemble learning and reinforcement learning, which fully considers the potential benefits and risks in different advertising delivery scenarios, realizes the optimal configuration of advertisers' budgets, avoids budget waste, and improves the return on investment of advertising delivery. At the same time, through multi-dimensional advertising delivery effect evaluation and real-time feedback mechanism, problems can be discovered in a timely manner and delivery strategies can be adjusted, and the advertising delivery effect can be continuously optimized, providing advertisers with an efficient, accurate, and intelligent advertising delivery solution. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 is a flow chart of the present invention; DETAILED DESCRIPTION

[0069] In the description of the present invention, it should be noted that the terms "front", "up", "down", "left", "right", "vertical", "horizontal", etc., indicating orientations or positional relationships, are based on the orientations or positional relationships shown in the accompanying drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, they cannot be understood as limiting the present invention.

[0070] refer to Figure 1 In order to solve the problems existing in the background technology, this application proposes the following technical solutions:

[0071] A method for generating an advertisement delivery engine comprises the following steps:

[0072] Step 1: First, collect multi-dimensional data of users through various channels, including but not limited to basic information of users (age, gender, region, etc.), browsing history, search history, purchase behavior, social data, etc. Suppose the collected user data set is U = {u1,u2,…,u n}, where n represents the number of users, u i Represents all data of the i-th user.

[0073] In step 1, the collected data is preprocessed, including data cleaning, missing value processing and normalization operations;

[0074] For missing values, multiple imputation method was used.

[0075] Assume that user feature x j If there are missing values, first build a prediction model (such as a random forest model) based on other complete features to predict the missing values, then generate m imputed data sets, conduct subsequent analysis on each imputed data set, and finally combine the m results to obtain the final conclusion;

[0076] Data normalization uses the minimum-maximum normalization method, and the formula is:

[0077]

[0078] where x ij represents the jth feature value of the i-th user, x′ ij Indicates the normalized eigenvalue, min(x j ) and max(x j ) represent the minimum and maximum values ​​of the jth feature, respectively. Data preprocessing can improve data quality and provide a reliable basis for subsequent analysis.

[0079] The technical solution is explained as follows: Regarding data collection, multi-dimensional user data is collected through a wide range of channels, covering basic information, browsing history, purchasing behavior, and more. This allows for a comprehensive user profile and deep insight into user preferences and needs. For example, not only is a user's age and gender known, but their search trends and spending habits are also understood. This allows advertising to be more closely aligned with the user's actual interests, significantly improving the accuracy of ad-user matching and ensuring that ads are no longer blindly pushed, but truly reach users with real needs.

[0080] Data cleaning removes noise and erroneous data, ensuring data authenticity and reliability and preventing contamination of analytical results by dirty data. Multiple imputation methods address missing values, using predictive models to generate multiple imputed data sets for comprehensive analysis. This maximizes data integrity and prevents analytical bias caused by missing information. Min-max normalization unifies data scales, making features of different magnitudes comparable and improving the efficiency and accuracy of subsequent model calculations. These combined operations lay a solid and reliable foundation for subsequent data-driven advertising strategy development and effectiveness evaluation, effectively advancing advertising delivery to a new level of precision, efficiency, and intelligence.

[0081] Step 2: In order to capture the user's dynamically changing interests in real time, a recurrent neural network (RNN) model based on the attention mechanism is used to model the user's dynamic interests;

[0082] In step 2, the specific method of using the recurrent neural network model based on the attention mechanism to model user dynamic interests is as follows:

[0083] First, modeling is performed based on the long short-term memory network. The specific steps are as follows:

[0084] Assume that the user's behavior sequence is S = {s1, s2, ..., s t}, where s t represents the user's behavior at time t;

[0085] The input gate formula of the long short-term memory network is: t =σ(W xi s t +W hi h t-1 +W ci c t-1 +b i ); the forget gate formula is: f t =σ(W xf s t +W hf h t-1 +W cf c t-1 +b f ); the output gate formula is: t =σ(W xo s t +W ho h t-1 +W co c t-1 +b o ); the cell state update formula is: c t =f t ⊙c t-1 +i t ⊙tanh(W xc s t +W hc h t-1 +b c ), the hidden state update formula is: h t =o t ⊙tanh(c t ); where σ is the Sigmoid function, W xi ,W hi ,W ci etc. is the weight matrix, b i ,b f ,b o ,b c is the bias vector, and ⊙ represents the element-wise product.

[0086] Introducing the attention mechanism, the formula for calculating the attention weight is: Where score(h i ,h j ) is used to calculate the hidden state h i and h j Functions that correlate between Through the attention mechanism, the model can focus on user behaviors that are more relevant to current interests and generate a more accurate user dynamic interest representation vector v.

[0087] The technical solution is explained as follows: Traditional methods struggle to adapt to the rapid changes in user interests in real time, while this model can track user behavior sequences in real time. For example, if a user's interest in travel products suddenly shifts to electronics within a short period of time, the LSTM can promptly adjust its judgment of user interests based on this new behavior sequence, ensuring that advertising delivery closely follows user interest dynamics.

[0088] The introduction of the attention mechanism enables the model to intelligently filter user behavior information and focus on the core components. For example, among a massive amount of user browsing and searching behavior, it can accurately identify the activities most relevant to current interests and discard irrelevant or minor information. For example, if a user repeatedly searches for details about a certain brand of mobile phone, the model will focus on these activities rather than other irrelevant browsing history.

[0089] The dynamic user interest representation vector generated in this way can more accurately reflect the user's current interests. This allows for highly relevant ads to be pushed to users based on more accurate interest vectors, significantly improving the appeal and relevance of ads, significantly increasing user attention and click-through rates, and delivering better results for advertisers, enhancing the relevance and effectiveness of advertising.

[0090] Step 3: Extract features from the advertising data, including content features (title, description, keywords, etc.), visual features (color and composition of pictures and videos, etc.), and delivery features (delivery platform, delivery time, target audience, etc.).

[0091] In step 3, for the text content features of the advertisement, a word vector model (such as Word2Vec) is used to map each word into a low-dimensional vector. Let the word set in the advertisement text be W = {w1, w2, ..., w m}, get the word vector e(w i ), and then the average pooling method is used to obtain the vector representation of the advertising text

[0092] For the visual features of advertisements, a convolutional neural network (CNN) is used to extract them. The method is as follows:

[0093] Let the advertisement image be I, and after the multi-layer convolution and pooling operation of the CNN network, the feature vector a is obtained. visual ;

[0094] Finally, the various feature vectors of the advertisement are concatenated to obtain the comprehensive feature representation vector of the advertisement:

[0095] a=[a text ;a visual ;a 投放特征 ].

[0096] The above technical solution is explained as follows: In content feature extraction, a word vector model (such as Word2Vec) is used to map the words in the advertising text into low-dimensional vectors and obtain a text vector representation through average pooling. This can accurately capture the semantic information of the advertising text and convert the product features and promotional highlights contained in the text into a vector form that can be understood by computers. For example, for a mobile phone advertisement, key semantics such as "high-performance processor" and "high-definition screen" can be clearly extracted, allowing the advertising delivery system to accurately understand the core content of the advertisement and provide strong support for subsequent precise matching of user needs.

[0097] Visual feature extraction uses convolutional neural networks (CNNs), which effectively extract information about visual elements such as color and composition from advertising images and videos. For example, a vibrantly colored and uniquely composed image of an electronic product advertisement can be precisely extracted by a CNN, enabling the system to perceive the ad's visual appeal.

[0098] By concatenating content, visual, and delivery feature vectors to create a comprehensive feature representation vector, we achieve the integration of all aspects of the ad. This provides the ad with a complete and accurate "digital profile" within the delivery system. When matching ads with users' dynamic interests, we can precisely connect them from multiple dimensions, greatly improving the fit between ads and user interests, making ad delivery more targeted and effective, thereby increasing ad display performance, click-through rates, and conversion rates, creating greater value for advertisers.

[0099] Step 4: Calculate the matching degree between users and ads by combining cosine similarity and deep matching network;

[0100] In step 4, based on the user dynamic interest representation vector υ and the advertisement comprehensive feature representation vector a, a method combining cosine similarity and deep matching network is used to calculate the user-advertisement matching method as follows:

[0101] The cosine similarity formula is: Where v·a represents the vector dot product, ||v|| and ||a|| represent the norms of vectors v and a respectively;

[0102] The deep matching network adopts a multi-layer perceptron (MLP) structure. The input is the concatenated vector [v; a], which is passed through multiple hidden layers h1, h2, ..., h l The output matching score is y deep ;

[0103] The hidden layer calculation formula is:

[0104] h k =σ(W k h k-1 +b k )(k=1,2,…,l);

[0105] The output layer formula is:

[0106] Final user-ad matching score score = α·sim cos (v,a)+(1-α)·y deep ;

[0107] Among them, α is the weight coefficient, which is determined through experimental optimization and is used to balance the contributions of cosine similarity and deep matching network, so as to more accurately measure the matching degree between users and ads.

[0108] The above technical solution is explained as follows: Cosine similarity measures the similarity between a user's dynamic interests and the comprehensive characteristics of an advertisement from a vector space perspective. It can quickly capture similarities between the two along basic feature dimensions. For example, if a user is interested in sports equipment and an advertisement features a sports brand, cosine similarity can quickly determine the degree of thematic compatibility between the two, providing a basic quantitative metric for matching calculations.

[0109] The Deep Matching Network, leveraging a multi-layer perceptron architecture, can deeply explore the complex, nonlinear relationships between users and ads. It can factor in the combined influence of multiple factors, such as user behavior and ad placement scenarios. For example, even for the same sports product ad, the Deep Matching Network can provide a more accurate match assessment based on these complex factors, targeting users with different purchase frequencies and viewing times.

[0110] Combining the two and determining the weighting coefficient α through experimental optimization can fully leverage the strengths of each. By leveraging cosine similarity to capture overall feature similarity and using a deep matching network to mine deeper connections, the matching score provides a more comprehensive and accurate measure of the degree of match between users and ads.

[0111] Step 5: Use ensemble learning methods to predict advertising effectiveness;

[0112] In step 5, the specific method of using ensemble learning method to predict the effect of advertising is as follows:

[0113] Combining multiple different prediction models, including gradient boosted decision trees (GBDT), support vector machines (SVM), and neural networks (NN);

[0114] Suppose the predicted click rate of user i seeing ad j by the kth prediction model is The final click-through rate prediction value is obtained by weighted average:

[0115] Where K is the number of models, w k is the weight of the k-th model, and w k The conversion rate prediction model is constructed by cross-validation optimization.

[0116] Suppose the behavior sequence after the user clicks the ad is B={b1,b2,…,b s}, using the sequence-to-sequence (Seq2Seq) model to predict the conversion rate r ij . By predicting click-through rate and conversion rate, a more comprehensive basis is provided for advertising placement decisions.

[0117] The above technical solution is explained as follows: Integrating multiple different models (such as GBDT, SVM, and NN) can combine the strengths of each model. Gradient boosting decision trees excel at handling nonlinear relationships and high-dimensional data, support vector machines perform well with small sample sizes, and neural networks have powerful feature learning capabilities. Combining these models can overcome the limitations of a single model and comprehensively capture factors related to advertising effectiveness. For example, when faced with complex user behavior data, different models can analyze from their respective strengths, and when combined, they can more accurately predict click-through rates.

[0118] The final click-through rate prediction is obtained through weighted averaging, and the weights are optimized through cross-validation, ensuring the accuracy and reliability of the prediction. Cross-validation can find the optimal weights for different data set partitions, ensuring that the model performs stably in various situations.

[0119] We also build a conversion rate prediction model that takes into account the user's behavioral sequence after clicking an ad. This allows predictions of ad effectiveness to move beyond click-through rate and encompass subsequent conversions, providing a more comprehensive basis for ad placement.

[0120] Step 6: Based on the predicted results of advertising effectiveness, a reinforcement learning method is used to dynamically allocate advertiser budgets;

[0121] In step 6, based on the predicted results of advertising effectiveness, the reinforcement learning method is used to dynamically allocate advertiser budgets. The specific method is as follows:

[0122] The advertising delivery process is modeled as a Markov decision process (MDP), where state s represents the current advertising delivery status, including the remaining budget, the effectiveness of the delivered ads, and the characteristics of the user group; action a represents the budget allocation strategy, such as the budget ratio allocated to different ad slots, time periods, or user groups; and reward r represents the benefits obtained after executing the action, such as the actual number of ad clicks, conversion rate, or revenue amount.

[0123] The Deep Q-Network (DQN) algorithm is used to learn the optimal budget allocation strategy. The Q-value function is expressed as Q(s, a; θ), where θ is a network parameter. The Q-value is updated according to the Bellman Equation by continuously interacting with the environment:

[0124]

[0125] Where α is the learning rate and γ is the discount factor. Through training, we can obtain the optimal budget allocation strategy, so that advertisers can obtain the maximum benefits under a given budget.

[0126] The above technical solution is explained as follows: Ad placement is modeled as a Markov decision process, comprehensively considering state factors such as remaining budget, ad effectiveness, and user group characteristics, as well as the action strategy for allocating budget to different ad placements, time periods, and user groups. Furthermore, actual revenue is used as a reward. This fully and accurately depicts the complex dynamics of ad placement. This enables budget allocation to move beyond simple empirical judgment or fixed patterns and instead be based on a deep understanding of the entire ad placement ecosystem.

[0127] The Deep Q Network algorithm continuously interacts with the environment to learn optimal budget allocation strategies. It dynamically adjusts budget allocation direction and proportions based on real-time changes in ad placement status and performance. For example, if a particular ad placement experiences a significant increase in conversion rate during a specific time period, the algorithm can detect this and promptly increase the budget for that placement and time period, capitalizing on the profit opportunity. Conversely, for underperforming placements, the algorithm can promptly reduce the budget to avoid wasted resources.

[0128] The optimal budget allocation strategy obtained through training can enable advertisers to maximize their profits under limited budgets.

[0129] Step 7: Based on the user and ad matching, ad delivery effect prediction and budget allocation results;

[0130] In step 7, based on the user and ad matching degree, ad delivery effect prediction, and budget allocation results, the ad delivery strategy is generated as follows:

[0131] First, the advertisements are sorted according to the matching degree and the predicted effect. Let the sorted advertisement list be A={a1,a2,…,a n};

[0132] Then, based on the budget allocation results, determine the budget, delivery time, and platform parameters for each ad.

[0133] For each user, select the ads with high matching degree and budget allowed from the sorted ad list for delivery;

[0134] Assume that user i’s available budget is budget i , the estimated cost of advertisement j is cost j , when budget i ≥cost j If the matching degree of ad j is above the threshold, ad j is delivered to user i and the available budget of user i is updated. i =budget i -cost j ;

[0135] The above technical solution is explained as follows: From the perspective of ad sorting, sorting by matching degree and predicted performance ensures that ads that are highly relevant to users' interests and have high predicted performance are prioritized. This ensures that the ads users see are more tailored to their needs, increasing their appeal and the likelihood of clicks and conversions. For example, for fitness enthusiasts, prioritizing ads related to fitness equipment and exercise classes with high predicted conversion rates can precisely meet user needs.

[0136] When determining advertising parameters, consider budget allocation results to set budget, time, and platform, ensuring optimal resource allocation. Avoid excessive advertising on inappropriate platforms or during inappropriate times, ensuring ads appear in scenarios where they are most effective. For example, during peak hours on fitness apps, you can target fitness enthusiasts with targeted fitness ads to maximize the effectiveness of ad exposure.

[0137] We select advertisements that are highly matched and within budget for each user, ensuring both the accuracy of advertisement delivery and cost control.

[0138] Step 8: Execute the advertising operation on the advertising platform according to the generated advertising strategy;

[0139] In step 8, according to the generated advertising strategy, the specific method for executing advertising delivery operations on the advertising delivery platform is as follows:

[0140] During the delivery process, user feedback data is collected in real time, including ad impressions, clicks, conversion rates, and user dwell time;

[0141] Let the collected impressions of ad j be impressions j , clicks j, the conversion rate is r ates j , the user stay time is time s pent j , store these data for subsequent effect evaluation and strategy optimization.

[0142] The technical solution described above is explained as follows: During ad placement, strictly adhering to the generated precision delivery strategy ensures that ads are displayed to target users at the appropriate time, platform, and with the appropriate budget and format. This effectively implements carefully designed strategies for matching and budget allocation, making ad placement more targeted and planned, and avoiding the waste of resources caused by blind advertising.

[0143] Collecting user feedback data in real time is crucial. Impressions reflect the reach of your ads, helping us understand their reach and assess the effectiveness of your delivery channels and timing. Clicks directly reflect a user's initial interest in your ad, allowing us to assess its appeal and relevance. Conversion rate is a key metric for measuring the ultimate effectiveness of an ad, demonstrating its ability to guide users to complete their intended actions (e.g., purchase, registration, etc.). User dwell time reflects the user's engagement and stickiness with the ad content.

[0144] Step 9: Use multi-dimensional evaluation indicators to evaluate the effectiveness of advertising;

[0145] In step 9: Use multi-dimensional evaluation indicators to evaluate the effectiveness of advertising. The specific method is as follows:

[0146] These include click-through rate (CTR), conversion rate (CVR), and return on investment (ROI). j , user engagement (such as user stay time, number of interactions, etc.);

[0147] The click-through rate formula is:

[0148] The conversion rate formula is:

[0149] The ROI formula is:

[0150] At the same time, the Entropy Weight Method is introduced to determine the weight of each evaluation indicator;

[0151] Let the jth index value of the i-th sample be x ij ; After normalization: The entropy value of the jth indicator Weight

[0152] Finally, the comprehensive evaluation score of advertisement j is obtained by combining the scores and weights of each indicator. Evaluate the effectiveness of advertising;

[0153] The above technical solution is explained as follows: Multi-dimensional evaluation metrics provide a comprehensive perspective. Click-through rate reflects the initial appeal of an ad, helping to determine whether the ad successfully captures user attention. Conversion rate measures the ad's ability to guide users to complete key behaviors, demonstrating its contribution to business goals. Return on investment (ROI) provides a direct view of advertising investment and output, allowing advertisers to clearly understand the effectiveness of their investment. User engagement metrics (such as dwell time and number of interactions) provide a deeper understanding of the depth and stickiness of user interaction with the ad. These metrics enable analysis of advertising effectiveness from different perspectives, avoiding the one-sidedness of a single metric.

[0154] The entropy weighting method determines indicator weights, objectively assigning values ​​based on the data's inherent discreteness. This avoids the potential bias introduced by subjective weighting, ensuring that each indicator's weighting more closely reflects the actual data characteristics. For example, in certain campaign scenarios, conversion rate may have a greater impact on overall effectiveness. The entropy weighting method can appropriately assign a higher weight based on data variations.

[0155] Step 10: Optimize the advertising delivery strategy based on the advertising effectiveness evaluation results.

[0156] In step 10, based on the evaluation results of the advertising delivery effect, the specific method for optimizing the advertising delivery strategy is as follows:

[0157] If an ad's overall evaluation score is low, analyze the reasons, which may be due to poor user-ad matching, unreasonable budget allocation, or inappropriate delivery time or platform. Adjust the corresponding parameters based on different reasons.

[0158] For example, if the issue is match quality, we can re-optimize user dynamic interest modeling and ad feature extraction methods, and update the user-ad match calculation model. If the issue is budget allocation, we can retrain the budget allocation strategy through reinforcement learning. If the issue is time or platform, we can adjust the time and platform settings for ad delivery. By continuously optimizing ad delivery strategies, we can achieve continuous improvement in ad delivery effectiveness.

[0159] In summary, in the present invention, at the user-advertising matching level, multi-dimensional data collection and preprocessing lay a solid foundation for accurate analysis. The matching degree calculation method based on LSTM of the attention mechanism and multi-model fusion can capture the dynamically changing interests of users in real time and accurately, deeply explore the complex potential relationship between users and advertisements, and break through the limitations of traditional methods that are difficult to adapt to behavioral changes and have weak ability to process high-dimensional sparse data. It greatly improves the accuracy of user-advertising matching, enhances the attractiveness and relevance of advertisements, and thus improves click-through rate and conversion rate.

[0160] In terms of budget allocation and effect prediction, ensemble learning combines multiple models to predict advertising effects, integrating the advantages of each model, comprehensively capturing influencing factors, and ensuring prediction accuracy; reinforcement learning models advertising as a Markov decision process, and realizes dynamic budget allocation through a deep Q-network algorithm. It flexibly adjusts strategies based on real-time delivery status and effects, avoids budget waste, achieves optimal allocation of advertisers' budgets, and improves return on investment.

[0161] To optimize the entire advertising process, we use multi-dimensional indicators combined with the entropy weight method to evaluate effectiveness, providing comprehensive and objective quantitative evaluations. Real-time feedback data is collected to inform strategic optimization. Based on these evaluations, we can precisely adjust targeting factors such as matchability, budget allocation, delivery timing, and platform. This creates a closed loop of "delivery-evaluation-optimization," continuously improving advertising effectiveness and creating efficient, intelligent, and precise advertising solutions for advertisers. This effectively enhances companies' market competitiveness and drives the advertising industry towards refined and intelligent development.

[0162] While the embodiments of the present invention have been shown and described, it will be apparent to those skilled in the art that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for generating an advertisement delivery engine, characterized in that: The specific steps include: Step 1: First, collect multi-dimensional data of users through various channels, including but not limited to basic information of users. Let the collected user data set be U = {u1,u2,…,u n }, where n represents the number of users, u i Represents all data of the i-th user; Step 2: Use a recurrent neural network model based on the attention mechanism to model user dynamic interests; Step 3: Extract features from the advertising data, including content features, visual features, and delivery features; Step 4: Calculate the matching degree between users and ads by combining cosine similarity and deep matching network; Step 5: Use ensemble learning methods to predict advertising effectiveness; Step 6: Based on the predicted results of advertising effectiveness, a reinforcement learning method is used to dynamically allocate advertiser budgets; Step 7: Based on the user and ad matching, ad delivery effect prediction and budget allocation results; Step 8: Execute the advertising operation on the advertising platform according to the generated advertising strategy; Step 9: Use multi-dimensional evaluation indicators to evaluate the effectiveness of advertising; Step 10: Optimize the advertising delivery strategy based on the advertising effectiveness evaluation results.

2. The method for generating an advertisement delivery engine according to claim 1, wherein: In step 1, the collected data is preprocessed, including data cleaning, missing value processing and normalization operations; For missing values, multiple imputation method is used as follows: Assume that user feature x j If there are missing values, first build a prediction model based on other complete features to predict the missing values, then generate m filled data sets, conduct subsequent analysis for each filled data set, and finally combine the m results to get the final conclusion; Data normalization uses the minimum-maximum normalization method, and the formula is: where x ij represents the jth feature value of the i-th user, x′ ij Indicates the normalized eigenvalue, min(x j ) and max(x j ) represent the minimum and maximum values ​​of the j-th feature respectively.

3. The method for generating an advertisement delivery engine according to claim 2, wherein: In step 2, the specific method of using the recurrent neural network model based on the attention mechanism to model user dynamic interests is as follows: First, modeling is performed based on the long short-term memory network. The specific steps are as follows: Assume that the user's behavior sequence is S = {s1, s2, ..., s t }, where s t represents the user's behavior at time t; The input gate formula of the long short-term memory network is: t =σ(W xi s t +W hi h t-1 +W ci c t-1 +b i ); the forget gate formula is: f t =σ(W xf s t +W hf h t-1 +W cf c t-1 +b f ); output The gate formula is: t =σ(W xo s t +W ho h t-1 +W co c t-1 +b o ); the cell state update formula is: c t =f t ⊙c t-1 +i t ⊙tanh(W xc s t +W hc h t-1 +b c ), the hidden state update formula is: h t =o t ⊙tanh(c t ); where σ is the Sigmoid function, W xi ,W hi ,W ci etc. is the weight matrix, b i ,b f ,b o ,b c is the bias vector, ⊙ represents the element-wise product; Introducing the attention mechanism, the formula for calculating the attention weight is: Where score(h i ,h j ) is used to calculate the hidden state h i and h j function of the correlation between them.

4. The method for generating an advertisement delivery engine according to claim 3, wherein: In step 3, for the text content features of the advertisement, the word vector model is used to map each word into a low-dimensional vector. Let the word set in the advertisement text be W = {w1, w2, ..., w m }, get the word vector e(w i ), and then the average pooling method is used to obtain the vector representation of the advertising text For the visual features of advertisements, a convolutional neural network is used to extract them. The method is as follows: Let the advertisement image be I, and after the multi-layer convolution and pooling operation of the CNN network, the feature vector a is obtained. visual ; Finally, the various feature vectors of the advertisement are concatenated to obtain the comprehensive feature representation vector of the advertisement: a=[a text ;a visual ;a 投放特征 ]。 5. The method for generating an advertisement delivery engine according to claim 4, wherein: In step 4, based on the user's dynamic interest representation vector υ and the advertisement comprehensive feature representation vector a, a method combining cosine similarity and deep matching network is used to calculate the user-advertisement matching degree as follows: The cosine similarity formula is: Where v·a represents the vector dot product, ||v|| and ||a|| represent the norms of vectors v and a respectively; The deep matching network adopts a multi-layer perceptron structure. The input is the concatenated vector [v; a], which passes through multiple hidden layers h1, h2, ..., h l The output matching score is y deep ; The hidden layer calculation formula is: h k =σ(W k h k-1 +b k )(k=1,2,…,l); The output layer formula is: Final user-ad matching score score = α·sim cos (v,a)+(1-α)·y deep ; Among them, α is the weight coefficient.

6. The method for generating an advertisement delivery engine according to claim 5, wherein: In step 5, the specific method of using ensemble learning method to predict the effect of advertising is as follows: Combine multiple different prediction models, including gradient boosted decision trees, support vector machines, and neural networks; Suppose the predicted click rate of user i seeing ad j by the kth prediction model is The final click-through rate prediction value is obtained by weighted average: Where K is the number of models, w k is the weight of the kth model, and w k Determined by cross-validation optimization; Suppose the behavior sequence after the user clicks the ad is B={b1,b2,…,b s }, using the sequence-to-sequence (Seq2Seq) model to predict the conversion rate r ij .

7. The method for generating an advertisement delivery engine according to claim 6, wherein: In step 6, based on the predicted results of advertising effectiveness, the reinforcement learning method is used to dynamically allocate advertiser budgets. The specific method is as follows: The advertising delivery process is modeled as a Markov decision process, where state s represents the current advertising delivery status, including the remaining budget, the effect of the delivered ads, and the characteristics of the user group. Action a represents the budget allocation strategy, and reward r represents the benefit obtained after executing the action; The deep Q-network algorithm is used to learn the optimal budget allocation strategy. The Q-value function is expressed as Q(s, a; θ), where θ is a network parameter. The Q-value is updated according to the Bellman equation by continuously interacting with the environment: Here, α is the learning rate and γ is the discount factor.

8. The method for generating an advertisement delivery engine according to claim 7, characterized in that: In step 7, based on the user and ad matching degree, ad delivery effect prediction, and budget allocation results, the ad delivery strategy is generated as follows: First, the advertisements are sorted according to the matching degree and the predicted effect. Let the sorted advertisement list be A={a1,a2,…,a n }; Then, based on the budget allocation results, determine the budget, delivery time, and platform parameters for each ad; For each user, select the ads with high matching degree and budget allowed from the sorted ad list for delivery; Assume that user i's budget is budget i , the estimated cost of advertisement j is cost j , when budget i ≥cost j When the matching degree of ad j is above the threshold, ad j is delivered to user i and the budget of user i is updated. i =budget i -cost j .

9. The method for generating an advertisement delivery engine according to claim 8, wherein: In step 8, according to the generated advertising strategy, the specific method for executing advertising delivery operations on the advertising delivery platform is as follows: During the delivery process, user feedback data is collected in real time, including ad impressions, clicks, conversion rates, and user dwell time; Let the collected impressions of ad j be impressions j , clicks j , the conversion rate is r ates j , the user stay time is time s pen.

10. The method for generating an advertisement delivery engine according to claim 9, wherein: In step 9: Use multi-dimensional evaluation indicators to evaluate the effectiveness of advertising. The specific method is as follows: These include click-through rate (CTR), conversion rate (CVR), and return on investment (ROI). j , user engagement; The click-through rate formula is: The conversion rate formula is: The ROI formula is: At the same time, the entropy weight method is introduced to determine the weight of each evaluation indicator; Let the jth index value of the i-th sample be x ij ; After normalization: The entropy value of the jth indicator Weight Finally, the comprehensive evaluation score of advertisement j is obtained by combining the scores and weights of each indicator. Index score j , evaluate the effectiveness of advertising.

Images