Advertisement design display system and method based on big data

Abstract

The invention discloses an advertisement design display system and method based on big data. The system comprises a data acquisition layer for acquiring multi-source data sources; the semantic understanding layer is used for generating semantic vectors of search words and advertising words based on a BERT-Chine pre-training language model, calculating a quantitative semantic matching degree through cosine similarity, performing deep semantic understanding and synonym equivalence matching, performing intention classification on features through an XGBoost gradient promotion decision tree, and judging that a user is in an instant purchase, decision period or interest exploration state; the scene sensing layer is used for dynamically recognizing a scene where a user is located by using an LSTM-CRF time sequence model; the feature recombination layer is used for extracting explicit features and mining implicit correlation features; the deep matching layer is used for generating a comprehensive matching score of a 0-1 interval by fusing explicit features and implicit association features in combination with multi-task learning, and performing screening according to the comprehensive matching score; the value measurement layer is used for quantifying contributions of different advertisement contacts to final conversion; and the execution layer is used for carrying out advertisement rendering and putting.

Description

Technical Field

[0001] This invention relates to the field of advertising design and display technology, and in particular to an advertising design and display system and method based on big data. Background Technology

[0002] In an era where the digital wave is sweeping the globe and information technology is advancing rapidly, big data-based advertising design and display, as the core carrier of data-driven marketing models, owes its birth and development to the increasing internet penetration, widespread mobile terminal penetration, and breakthroughs in artificial intelligence technology, which have enhanced multi-dimensional data collection and analysis capabilities. By integrating massive amounts of heterogeneous information such as user behavior patterns, consumption preferences, social interactions, and geographical locations, it constructs precise user profiles and uses machine learning algorithms to predict demand trends. This not only breaks down the information barriers of traditional extensive advertising placement, achieving personalized matching and dynamic optimization of advertising content and target audiences, but also fundamentally reconstructs the value transmission chain between advertisers and consumers. Through the effective conversion of data assets, it improves advertising reach efficiency and conversion rates, reduces marketing costs, and enhances the immersiveness and resonance of user experience. This, in turn, drives the advertising industry to transform from one-way communication to two-way interaction, injecting strong momentum into the innovation, upgrading, and sustainable development of the advertising industry while promoting the deep integration of the digital economy and the real economy.

[0003] In existing technologies, the matching of user search keywords with advertising templates is limited to surface word frequency statistics and does not establish a semantic association network. For example, a user searching for "outdoor equipment" may associate it with trekking poles rather than sunscreen, but the system may miss it due to the low keyword repetition rate. Therefore, an advertising design and display system and method based on big data is proposed. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing an advertising design and display system and method based on big data.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: An advertising design and display system based on big data, comprising: Data Acquisition Layer: Integrates multiple data sources including device sensors, network status monitoring, and third-party APIs. Employs Kafka real-time stream processing and Spark structured computing technologies to perform high-concurrency acquisition and preprocessing of user behavior logs, spatiotemporal environmental data, and social media buzz. Semantic understanding layer: Based on the BERT-Chinese pre-trained language model, semantic vectors of search terms and advertising terms are generated. The semantic matching degree is quantified by calculating cosine similarity. Deep semantic understanding and synonym equivalence matching are performed. The XGBoost gradient boosting decision tree is used to classify the intent of the features and determine whether the user is in an immediate purchase, decision-making period or interest exploration state. Scene perception layer: It integrates GPS speed, network type and time series data, uses LSTM-CRF time series model to dynamically identify the scene in which the user is located, and combines environmental sensor data to optimize the scene label accuracy; Feature Recombination Layer: Combines user profiles, spatiotemporal context, and behavioral sequences using Cartesian product and DeepFM automatic feature cross-interaction to extract explicit features and uncover implicit correlation features; Deep matching layer: By fusing explicit features and implicit association features with multi-task learning, it simultaneously optimizes three objectives: semantic matching degree, cross feature score and user intent weight, and generates a comprehensive matching score in the range of 0-1. The selection is then based on the comprehensive matching score. Value Measurement Layer: Through dynamic weight allocation and Shapley Value attribution analysis, the contribution of different advertising touchpoints to the final conversion is quantified; Execution layer: Based on the max-heap priority queue and dynamic frequency control algorithm, combined with the user lifecycle value stratification strategy, ad rendering and delivery are performed.

[0006] The above technical solution further includes: Preferably, the specific steps for the semantic understanding layer to classify customer intent are as follows: Semantic vector generation: The input text is preprocessed, Chinese text is segmented by character, and [CLS] and [SEP] tags are added to the beginning and end of the text. Characters are converted into corresponding IDs in the BERT vocabulary. BERT generates context-aware vectors through a 12-layer bidirectional Transformer encoder for the input sequence. The hidden layer output of the i-th character Represented as , To exclude the context sequence of the i-th character, a total semantic vector is generated, represented as follows: ; Cosine similarity calculation: Calculate the cosine similarity between user search term U and advertising keyword A, expressed as... , For users to search for word vectors, For advertising keyword vectors, For the user's search term vector length, The vector length of the advertising keywords; Synonym equivalence matching: Synonym clustering is achieved by calculating the Euclidean distance between words in the semantic space; Feature engineering: Constructing a vector space through feature engineering; XGBoost predicts probabilities: The XGBoost gradient boosting decision tree model is used to iteratively build a decision tree by fitting the residuals of the previous trees. The objective function is to minimize the residuals containing the first-order gradient. Second gradient and regularization term loss function Controlling model complexity; Classification decision: Intent labels are assigned based on predicted probabilities. When the probability is greater than 0.7, it is judged as an immediate purchase; between 0.4 and 0.7 is the decision period; and less than or equal to 0.4 is the interest exploration period.

[0007] Preferably, the specific steps for the scene perception layer to identify the user's scene are as follows: Temporal feature engineering: Constructing a multi-dimensional feature vector containing normalized velocity values, network encoding, time periodic features, and normalized temperature and humidity values; Employing a sliding window strategy to process non-uniform length temporal data and capture contextual dependencies. LSTM-CRF model: A hybrid model based on LSTM gating mechanism and CRF sequence optimization for dynamic scene recognition, where LSTM handles long-term dependencies and CRF corrects local decision contradictions; Environmental data optimization: Environmental sensor data is introduced and the initial scene probability is fused with environmental evidence through a Bayesian update framework, and a dynamic weight adjustment strategy is adopted to optimize the final scene label accuracy.

[0008] Preferably, the specific steps of the feature reorganization layer in performing high-order feature interaction through feature engineering and deep learning are as follows: Cartesian product explicit combination: Using the mathematical Cartesian product operation, the user profile, spatiotemporal context, and behavioral sequence feature domains are fully cross-combined to generate explicit combination features of order three or higher; DeepFM Implicit Cross Modeling: Through a joint architecture of factorization machine and deep neural network, the factorization machine part uses vector dot product to calculate second-order feature crosses to capture the linear relationship between features, while the deep neural network part learns higher-order feature interactions through multi-layer nonlinear transformations to discover deep, nonlinear implicit association patterns and generate implicit cross features with business interpretability.

[0009] Preferably, the specific steps for filtering advertisements using the deep matching layer are as follows: Feature Input and Forward Propagation in Deep Neural Networks: Explicit Feature Vectors With implicit feature vectors Concatenate into a joint feature vector The input-degree neural network model consists of L fully connected layers. The first L−1 layers use the ReLU activation function, and the last layer outputs the predicted values ​​of the three nodes. The output is constrained to the [0,1] interval by the Sigmoid function. The forward propagation process is represented as follows: ,in, For the Sigmoid function, This is the output of the l-th layer. Let be the weight matrix of the l-th layer. Let l be the bias vector of the l-th layer. This is the predicted value for semantic matching degree. The predicted value is the cross-feature score. The predicted value for user intent weight; Multi-task learning: Simultaneously optimize three objectives—semantic matching degree, cross-feature score, and user intent weight—through multi-task learning, and generate a comprehensive matching score through a weighted fusion strategy; The selection strategy based on comprehensive scores is as follows: the scores are compressed to the 0-1 range using the Sigmoid function, and the ads are selected according to the preset threshold. Scores greater than 0.8 are directly entered into the candidate pool, scores between 0.5 and 0.8 need to be sorted again based on scene adaptability, and scores below 0.5 are eliminated.

[0010] Preferably, the specific steps of the value measurement layer to quantify the contribution of different advertising touchpoints to the final conversion are as follows: Weighting Calculation: Let the ad type be T∈{Brand Ads, Performance Ads}, and the overall value score be calculated. for ,in, For the growth rate of advertising search volume, For social media buzz, LTV is user lifetime value, CTR is click-through rate, and CVR is conversion rate; Secondary ranking strategy: When the comprehensive value score When ∈[0.5,0.8], a secondary sort is performed based on scene adaptability: , The scene weight is used for scene adaptation, which is the degree to which the ad content matches the current scene. Shapley Value Attribution Analysis: This analysis quantifies the contribution of different ad touchpoints to the final conversion using Shapley Value. Let the set of ad touchpoints be N = {1, 2, ..., n}, and the Shapley Value of touchpoint i be... for , The marginal contribution of the contact subset S, The number of combinations represents the size of all possible subsets. If the Shapley Value of search ads exceeds 50%, the search ad budget will be increased, and the landing page conversion rate will be optimized. If the Shapley Value of display ads is less than 20%, but the CTR is artificially high, the display ad bid will be reduced, and the creative relevance will be optimized.

[0011] Preferably, the specific steps for ad rendering and delivery in the execution layer are as follows: Max-heap priority queue construction: Ads are constructed into a max-heap based on their overall matching scores. The max-heap satisfies the condition that the overall matching score is greater than or equal to the overall score of each ad. When a new ad enters the queue or the score of an existing ad is updated, the heap structure is maintained through an up-filtering operation. Dynamic frequency control: Let the LTV of user u be... The frequency control parameters for advertisement a are: ,in, For LTV stratification threshold, Based on the upper limit of basic frequency control, To set the maximum frequency control limit, record the number of times ad a is displayed for user u within the last T days, C(u,a). If so, display is allowed; otherwise, skip. The time decay factor, The time since the most recent display; LTV layered display: Greater than or equal to Indicating high value, the frequency control upper limit f(u,a) is: The advertising content preference is personalized recommendations and brand advertising; exist and The value between them is represented by the frequency control upper limit f(u,a) as follows: The advertising content preference is for category-related ads and light promotions. Less than or equal to To indicate low value, the frequency control upper limit f(u,a) is... The advertising content preferences are high-discount ads and new user gift packs.

[0012] A big data-based advertising design and display method includes the following steps; Multi-source data acquisition and preprocessing: Integrating multiple data sources, using Kafka real-time stream processing technology to build a high-concurrency data pipeline to buffer and distribute user behavior logs, spatiotemporal environmental data and social media volume, and combining the Spark structured computing engine to complete data cleaning, format standardization and lightweight feature extraction; Deep User Intent Analysis: Based on the BERT-Chinese pre-trained language model, user search terms and advertising copy are transformed into high-dimensional semantic vectors. The semantic matching degree between the two is quantified by cosine similarity calculation. At the same time, XGBoost gradient boosting decision tree is used to classify user behavior features and provide dynamic weight coefficients for the deep matching layer. Dynamic scene perception: By integrating GPS speed, network type and time series data, an LSTM-CRF time series model is constructed to dynamically identify the user's scene, and the scene label accuracy is further optimized by combining environmental sensor data; Intelligent feature recombination: User profiles, spatiotemporal context, and behavioral sequences are combined using Cartesian products to generate explicit features, and then combined with the DeepFM automatic feature cross-model to mine implicit related features; Deep matching and comprehensive score generation: Construct a DNN multi-task learning model, integrate explicit features and implicit association features, and optimize semantic matching degree, cross feature score (DeepFM output) and user intent weight to generate a comprehensive matching score in the range of 0-1. Ads with a score > 0.8 directly enter the candidate pool, and ads with a score of 0.5-0.8 are sorted in a second time based on scene adaptability. Multi-objective value assessment: By collecting real-time feedback metrics through front-end tracking, brand value metrics from search engine data / social media APIs, user quality metrics from order data / CRM systems, and negative feedback metrics from user behavior logs, a multi-objective evaluation indicator system is constructed. Combined with dynamic weight allocation and Shapley Value attribution analysis, the contribution of different advertising touchpoints to the final conversion is quantified. Ad design and display: A max-heap priority queue is constructed based on the comprehensive matching score to ensure that high-value ads are displayed first. Dynamic frequency control is implemented in combination with the user LTV stratification strategy, and dynamic ad templates are generated according to the scenario adaptation strategy.

[0013] The present invention has the following beneficial effects: 9. In this invention, a pre-trained language model is introduced to upgrade the matching of user search keywords and advertising template keywords from "word frequency counting" to "semantic vector similarity calculation". This captures the deep semantic relationship between keywords, especially distinguishing the different intentions of similar users and avoiding interference from low-value ads to high-value users. By combining semantic vectors and intent weights, the matching of advertising templates and user needs is upgraded from "keyword coverage" to "needs understanding", thereby improving the conversion rate.

[0014] 10. In this invention, GPS speed, network type and time series data are integrated to construct an LSTM-CRF time series model to dynamically identify the user's scene, and scene labels are further optimized by combining environmental sensor data. Advertising strategies are customized according to scene labels to avoid advertising waste caused by scene mismatch. Attached Figure Description

[0015] Figure 1 This is a system architecture diagram of an advertising design and display system based on big data proposed in this invention; Figure 2 This is a flowchart of an advertising design and display method based on big data proposed in this invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] like Figures 1-2 As shown, an advertising design and display system based on big data includes: Data Acquisition Layer: Integrates multiple data sources including device sensors, network status monitoring, and third-party APIs. Employs Kafka real-time stream processing and Spark structured computing technologies to perform high-concurrency acquisition and preprocessing of user behavior logs, spatiotemporal environmental data, and social media buzz. Semantic understanding layer: Based on the BERT-Chinese pre-trained language model, semantic vectors of search terms and advertising terms are generated. The semantic matching degree is quantified by cosine similarity calculation, which breaks through the superficial association of traditional keyword matching. Deep semantic understanding and synonym equivalence matching are used. The XGBoost gradient boosting decision tree is used to classify intent based on 200+ features such as click-through rate comparison behavior and shopping cart abandonment rate, and to determine whether the user is in an immediate purchase, decision-making period or interest exploration state, providing dynamic weight coefficients for the DNN matching model. Scene perception layer: It integrates GPS speed, network type and time series data, uses LSTM-CRF time series model to dynamically identify the user's scene (such as commuting / home / office), and combines environmental sensor data (temperature, humidity / light) to optimize scene label accuracy; Scene adaptation strategy: During commuting (while on the move), push short, efficient ads (short videos / large fonts); At home, push long-form content ads (product reviews / special offers); For office use, block entertainment ads and push office supplies; Feature Recombination Layer (Cross Feature Engine): Combines user profile (age / gender), spatiotemporal context (time period / geographical location), and behavioral sequence (recently viewed categories) with Cartesian product and DeepFM automatic feature cross-extraction to extract explicit features and uncover implicit association features such as "25-year-old female + weekend afternoon + shopping mall". Deep matching layer (Deep Neural Network (DNN) matching model): By fusing explicit features and implicit association features with multi-task learning, it simultaneously optimizes three objectives: semantic matching degree, cross feature score, and user intent weight, generating a comprehensive matching score in the range of 0-1, and then filtering based on the comprehensive matching score; Ads with a score > 0.8 are directly entered into the candidate pool; Ads scoring 0.5-0.8 need to be re-ranked based on their suitability for the given context. Value Measurement Layer (Multi-Objective Evaluation System): Through dynamic weight allocation (ads with scores of 0.5-0.8 need to be re-ranked based on scenario adaptability, i.e., brand ads focus on social media presence, while performance ads focus on LTV) and Shapley Value attribution analysis, the contribution of different ad touchpoints to the final conversion is quantified, achieving a leap from a single CTR indicator to comprehensive value evaluation and avoiding misleading strategies with inflated CTR. Multi-objective evaluation index system: Indicator type, measurement dimension, and data collection method Real-time feedback on CTR (Click-Through Rate), CVR (Conversion Rate), and dwell time via front-end tracking. Brand value, brand search volume growth rate, social media buzz (search engine data / social media API) User quality conversion, user lifetime value (LTV), order data / CRM system Negative feedback ad close rate, complaint tags, user behavior logs Dynamic weight allocation: Brand advertising: Weighting = 60% brand search volume + 30% social media buzz + 10% CTR; Performance advertising: Weighting = 50% CVR + 30% LTV + 20% CTR; Execution layer (display decision engine): Based on the max-heap priority queue and dynamic frequency control algorithm, combined with the user lifetime value (LTV) tiering strategy, it performs ad rendering and delivery.

[0018] In this embodiment: Raw data acquisition: Device sensors → Data acquisition module (Kafka real-time stream) Feature preprocessing: Text data → Semantic understanding layer (BERT vectorization); Spatiotemporal data → Scene awareness layer (LSTM modeling); Behavioral sequence → Semantic understanding layer (XGBoost classification); Feature fusion: semantic features + scene features + intent features → cross-feature engine (DeepFM combination); Deep matching: cross features → DNN model → generate matching score (0-1); Value assessment: Matching score + multi-dimensional indicators → comprehensive evaluation model → calculation of ad value; Decision execution: Evaluation results → Display engine → Trigger ad rendering and delivery.

[0019] Scenario 1: On-demand purchasing needs during commuting User A (28-year-old male) searched for "Bluetooth headset" on the subway; The semantic understanding layer identifies advertisements related to "noise-canceling headphones" and determines them to be in the "decision period" (multiple price comparisons in the past 3 days). The scene perception layer determines the scene as a commuting scenario based on GPS speed (>30km / h) and network type (mobile data); The feature recombination layer combines the features of "male + commuting + audio device", and the DNN model gives a matching score of 0.89; The value measurement layer assigns this ad a high priority (CVR prediction value 15%). The display engine pushes "30-second unboxing video" ads (vertical full screen, with a quick purchase button); Scenario 2: Brand Exposure in the Home Environment User B (35-year-old female) browsed "Maternity and Baby Products" on a weekend afternoon. The semantic understanding layer matches the "organic milk powder" advertisement and determines it as "interest exploration" (first time browsing this category); The scene perception layer determines the scene as a home based on Wi-Fi connection and device orientation (landscape). The feature recombination layer combined the features of "female + weekend + maternal and infant", and the matching score was 0.78; The value measurement layer assigns high weight to brand advertising (20% predicted growth in social media buzz); The display engine pushes a long video ad for "Expert Parenting Lecture" (horizontal immersive view with brand logo watermark).

[0020] In one embodiment, the specific steps for the semantic understanding layer to classify customer intent are as follows: Semantic vector generation: The input text is preprocessed, Chinese text is segmented by character (BERT-Chinese uses character-level granularity), [CLS] and [SEP] tags are added to the beginning and end of the text, and characters are converted into corresponding IDs in the BERT vocabulary. BERT generates context-aware vectors through a 12-layer bidirectional Transformer encoder for the input sequence. The hidden layer output of the i-th character Represented as , To exclude the context sequence of the i-th character, a total semantic vector is generated, represented as follows: ; Cosine similarity calculation: Calculate the cosine similarity between user search term U and advertising keyword A, expressed as... , For users to search for word vectors, For advertising keyword vectors, For the user's search term vector length, The vector length of the advertising keywords; Synonym equivalence matching: Synonym clustering is achieved by calculating the Euclidean distance between words in the semantic space; Feature Engineering: A vector space with 200+ dimensions of features is constructed through feature engineering, including click-through price comparison behavior, shopping cart abandonment rate, page dwell time, and recent purchase interval. Among them, the price comparison behavior feature is defined as the ratio of the number of price comparison clicks to the total number of clicks, the abandonment rate feature is defined as the ratio of the number of times the cart is added minus the number of purchases to the number of times the cart is added, the dwell time feature is defined as the average dwell time of all pages, and the purchase interval feature is defined as the difference between the current time and the last purchase time. XGBoost predicts probabilities: The XGBoost gradient boosting decision tree model is used to iteratively build a decision tree by fitting the residuals of the previous trees. The objective function is to minimize the residuals containing the first-order gradient. Second gradient and regularization term loss function Controlling model complexity; Classification decision: Intent labels are assigned based on predicted probabilities. When the probability is greater than 0.7, it is judged as an immediate purchase; between 0.4 and 0.7 is the decision period; and less than or equal to 0.4 is the interest exploration period.

[0021] In one embodiment, the specific steps of the scene perception layer in identifying the user's scene are as follows: Temporal feature engineering: Construct a multi-dimensional feature vector containing normalized velocity values, network encoding, time periodic features (using sine and cosine encoding to process hourly periodicity), and normalized temperature and humidity values; Employ a sliding window strategy to process non-uniform length time-series data and capture contextual dependencies. LSTM-CRF model: A hybrid model for dynamic scene recognition based on LSTM gating mechanism (forget gate controls the retention of historical information, input gate regulates the writing of new information, and output gate determines the update of hidden state) and CRF sequence optimization (constraining the legality of label transfer through conditional probability model). LSTM handles long-term dependencies, and CRF corrects local decision contradictions. Environmental data optimization: Environmental sensor data is introduced and fused with initial scene probabilities and environmental evidence (such as reducing confidence in commuting scenarios when temperatures are abnormal) using a Bayesian update framework. A dynamic weight adjustment strategy is then employed (assigning different weights to environmental data based on the degree of temperature and humidity deviation) to optimize the final scene label accuracy. Feature reorganization layer (cross-feature engine): User profiles (age / gender), spatiotemporal context (time period / geographical location), and behavioral sequences (recently viewed categories) are combined using Cartesian products and automatically cross-referenced with DeepFM to uncover implicit association features such as "25-year-old female + weekend afternoon + shopping mall".

[0022] In one embodiment, the specific steps of the feature reorganization layer in performing high-order feature interaction through feature engineering and deep learning are as follows: Cartesian product explicit combination: Using the Cartesian product operation in mathematics, the feature domains of user profile (age / gender), spatiotemporal context (time period / geographical location), and behavioral sequence (recently viewed categories) are fully cross-combined to generate explicit combination features of level three or above, such as "25-year-old female + weekend afternoon + shopping mall". This technique expands the feature space by exhaustively enumerating all possible feature combinations and effectively captures the explicit correlation between user attributes, environmental factors and behavioral patterns. DeepFM Implicit Cross Modeling: This approach utilizes a joint architecture of Factorization Machine (FM) and Deep Neural Network (DNN). The FM component uses vector dot products to calculate second-order feature interactions (such as the interaction between age and time period, or gender and geographic location), capturing linear relationships between features. The DNN component learns higher-order feature interactions (such as the fifth-order nonlinear association between age, gender, time period, geographic location, and product category) through multi-layer nonlinear transformations, discovering deep, nonlinear implicit association patterns. By automatically learning feature interaction weights, it avoids the complexity of manual feature engineering. It combines the linear interpretability of FM with the nonlinear expressive power of DNN to generate business-interpretive implicit cross features, such as "a 25-year-old woman browses cosmetics in a shopping mall on a weekend afternoon." These explicit and implicit features are input into the DNN matching model, significantly improving the accuracy of ad matching. Explicit combined features address the problem of feature omissions in traditional feature engineering, while implicit cross features solve the difficulty of manually constructing higher-order associations. Their synergistic effect enhances feature expressive power from basic features (71% AUC) to combined features (78%). From AUC to implicit cross features (85% AUC), it provides high-order, non-linear feature representations for advertising recommendation systems, supporting accurate personalized advertising matching decisions.

[0023] In one embodiment, the specific steps for filtering advertisements using the deep matching layer are as follows: Feature Input and Forward Propagation of Deep Neural Networks (DNNs): Explicit Feature Vectors (Such as user age, gender, and advertising category tags) and implicit feature vectors (Such as embedding user historical click sequences, semantic embedding of ad images / text) are concatenated into a joint feature vector. The input degree neural network (DNN) model consists of L fully connected layers. The first L−1 layers use the ReLU activation function, and the last layer outputs the predicted values ​​of three nodes (corresponding to semantic matching degree, cross feature score, and user intent weight, respectively). The output is restricted to the [0,1] interval by the Sigmoid function. The forward propagation process is represented as follows. ,in, For the Sigmoid function ( ), This is the output of the l-th layer. Let be the weight matrix of the l-th layer. Let l be the bias vector of the l-th layer. This is the predicted value for semantic matching degree. The predicted value is the cross-feature score. The predicted value for user intent weight; Multi-task learning: Simultaneously optimize semantic matching through multi-task learning (using the cross-entropy loss function to quantify the difference between predicted and true values). Cross-feature score (measured by the mean squared error (MSE) loss function) User intent weights (adjusting the prediction bias of intent labels using the mean squared error loss function) The three objectives are to be achieved through a weighted fusion strategy (dynamically allocating weight coefficients for semantics, cross-references, and intent). Generate a comprehensive matching score; The selection strategy based on comprehensive scores is as follows: the scores are compressed to the 0-1 range using the Sigmoid function, and the ads are selected according to the preset threshold. Scores greater than 0.8 are directly entered into the candidate pool, scores between 0.5 and 0.8 need to be sorted again based on scene adaptability, and scores below 0.5 are eliminated.

[0024] In one embodiment, the value measurement layer quantifies the specific steps of different ad touchpoints contributing to the final conversion: Weighting Calculation: Let the ad type be T∈{Brand Ads, Performance Ads}, and the overall value score be calculated. for ,in, For the growth rate of advertising search volume, For social media buzz, LTV is the lifetime value of a user (calculated through a CRM system), CTR is the click-through rate (number of ad clicks ÷ number of ad impressions × 100%), and CVR is the conversion rate (number of users who completed the target action ÷ number of users who clicked the ad × 100%). Secondary ranking strategy: When the comprehensive value score When ∈[0.5,0.8], a secondary sort is performed based on scene adaptability: , The scene weight is used for scene adaptation, which is the degree to which the ad content matches the current scene. Shapley Value Attribution Analysis: This analysis quantifies the contribution of different ad touchpoints (such as search ads, display ads, and email marketing) to the final conversion using Shapley Value. Let the set of ad touchpoints be N = {1, 2, ..., n}, and the Shapley Value of touchpoint i be... for , Marginal contribution of the touchpoint subset S (e.g., conversion rate after the subset S reaches the user). The number of combinations represents the size of all possible subsets. If the Shapley Value of search ads accounts for more than 50%, the search ad budget will be increased, and the landing page conversion rate will be optimized. If the Shapley Value of display ads is less than 20%, but the CTR is artificially high (e.g., a click-through rate of 3% but a conversion rate of 0.5%), the display ad bid will be reduced, and the relevance of the creative will be optimized.

[0025] In one embodiment, the specific steps for the execution layer to render and deliver advertisements are as follows: Max-heap priority queue construction: Ads are constructed into a max-heap based on their overall matching scores to ensure that high-priority ads are displayed first. The max-heap satisfies that the overall matching score is greater than or equal to the overall score of each ad. When a new ad enters the queue or the score of an existing ad is updated, the heap structure is maintained through an up-filtering operation. Dynamic frequency control: Let the LTV of user u be... The frequency control parameters for advertisement a are: ,in, For LTV stratification threshold, Based on the upper limit of basic frequency control, To set the maximum frequency control limit, record the number of times ad a is displayed for user u within the last T days, C(u,a). If so, display is allowed; otherwise, skip. The time decay factor, The time since the most recent display; LTV layered display: Greater than or equal to Indicating high value, the frequency control upper limit f(u,a) is: The advertising content preference is personalized recommendations and brand advertising; exist and The value between them is represented by the frequency control upper limit f(u,a) as follows: The advertising content preference is for category-related ads and light promotions. Less than or equal to To indicate low value, the frequency control upper limit f(u,a) is... The advertising content preferences are high-discount ads and new user gift packs.

[0026] A big data-based advertising design and display method includes the following steps; Multi-source data acquisition and preprocessing: Integrating multiple data sources such as device sensors, network status monitoring, and third-party APIs, using Kafka real-time stream processing technology to build a high-concurrency data pipeline to buffer and distribute user behavior logs, spatiotemporal environmental data, and social media buzz, and combining the Spark structured computing engine to complete data cleaning (duplicate removal, outlier handling), format standardization (timestamp unification), and lightweight feature extraction (such as statistics on recently viewed categories by users); Deep analysis of user intent: Based on the BERT-Chinese pre-trained language model, user search terms and advertising copy are transformed into high-dimensional semantic vectors. The semantic matching degree between the two is quantified by cosine similarity calculation. At the same time, XGBoost gradient boosting decision tree is used to classify intent of 200+ dimensions of user behavior features such as click comparison behavior and shopping cart abandonment rate (determining whether the user is in an immediate purchase, decision-making period or interest exploration state), and dynamic weight coefficients are provided for the deep matching layer. Dynamic Scene Awareness: By integrating GPS speed (to determine movement / stationary status), network type (4G / Wi-Fi), and time series data, an LSTM-CRF time series model is constructed to dynamically identify the user's scene (such as commuting, home, and office). Environmental sensor data (temperature, humidity, light intensity) is combined to further optimize the accuracy of scene labels (such as refining the "home-night-low temperature" scene). The scene adaptation strategy executes differentiated push based on the scene awareness results. For example, in commuting scenarios (while moving), short and efficient ads such as 15-second videos and large-font text are pushed to adapt to fragmented attention. In home scenarios, long-content ads such as product review articles and discount combination details pages are pushed to meet the needs of in-depth decision-making. In office scenarios, entertainment ads are blocked and office supplies such as printers and stationery are prioritized to avoid interfering with the user's work state. This achieves the function of precise adaptation between scene and ad format. Intelligent feature recombination: The Cartesian product of user profiles (age, gender, spending power), spatiotemporal context (time period, geographical location) and behavioral sequence (recently viewed categories) is used to generate explicit features, and the DeepFM automatic feature cross-model is used to mine implicit related features. Deep matching and comprehensive score generation: Construct a DNN multi-task learning model, integrate explicit features and implicit association features, and simultaneously optimize three objectives: semantic matching degree (BERT vector similarity), cross feature score (DeepFM output), and user intent weight (XGBoost classification result). Generate a comprehensive matching score in the range of 0-1, where ads with a score >0.8 directly enter the candidate pool, and ads with a score of 0.5-0.8 are sorted a second time based on scene adaptability. Multi-objective value assessment: Real-time feedback metrics (CTR, CVR, dwell time) are collected through front-end tracking, brand value metrics (brand search volume growth rate, social media buzz) are collected through search engine data / social media APIs, user quality metrics (converted user LTV) are collected through order data / CRM system, and negative feedback metrics (ad close rate, complaint tags) are collected through user behavior logs. A multi-objective evaluation indicator system is constructed, and dynamic weight allocation and Shapley Value attribution analysis are combined to quantify the contribution of different ad touchpoints to the final conversion. Ad design and display: Based on the comprehensive matching score, a max-heap priority queue is constructed to ensure that high-value ads are displayed first. Combined with the user LTV stratification strategy, dynamic frequency control is implemented (reducing the frequency of high-value users to improve the experience, and increasing the frequency of new users to accelerate conversion). Dynamic ad templates are generated according to the scenario adaptation strategy (such as automatically cropping to short video format for commuting scenarios).

[0027] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An advertising design and display system based on big data, characterized in that, include: Data Acquisition Layer: Integrates multiple data sources including device sensors, network status monitoring, and third-party APIs. Employs Kafka real-time stream processing and Spark structured computing technologies to perform high-concurrency acquisition and preprocessing of user behavior logs, spatiotemporal environmental data, and social media buzz. Semantic understanding layer: Based on the BERT-Chinese pre-trained language model, semantic vectors of search terms and advertising terms are generated. The semantic matching degree is quantified by calculating cosine similarity. Deep semantic understanding and synonym equivalence matching are performed. The XGBoost gradient boosting decision tree is used to classify the intent of the features and determine whether the user is in an immediate purchase, decision-making period or interest exploration state. Scene perception layer: It integrates GPS speed, network type and time series data, uses LSTM-CRF time series model to dynamically identify the scene in which the user is located, and combines environmental sensor data to optimize the scene label accuracy; Feature Recombination Layer: Combines user profiles, spatiotemporal context, and behavioral sequences using Cartesian product and DeepFM automatic feature cross-interaction to extract explicit features and uncover implicit correlation features; Deep matching layer: By fusing explicit features and implicit association features with multi-task learning, it simultaneously optimizes three objectives: semantic matching degree, cross feature score and user intent weight, and generates a comprehensive matching score in the range of 0-1. The selection is then based on the comprehensive matching score. Value Measurement Layer: Through dynamic weight allocation and Shapley Value attribution analysis, the contribution of different advertising touchpoints to the final conversion is quantified; Execution layer: Based on the max-heap priority queue and dynamic frequency control algorithm, combined with the user lifecycle value stratification strategy, ad rendering and delivery are performed.

2. The advertising design and display system based on big data according to claim 1, characterized in that, The specific steps for the semantic understanding layer to classify customer intent are as follows: Semantic vector generation: The input text is preprocessed, Chinese text is segmented by character, and [CLS] and [SEP] tags are added to the beginning and end of the text. Characters are converted into corresponding IDs in the BERT vocabulary. BERT generates context-aware vectors through a 12-layer bidirectional Transformer encoder for the input sequence. The hidden layer output of the i-th character Represented as , To exclude the context sequence of the i-th character, a total semantic vector is generated, represented as follows: ; Cosine similarity calculation: Calculate the cosine similarity between user search term U and advertising keyword A, expressed as... , For users to search for word vectors, For advertising keyword vectors, For the user's search term vector length, The vector length of the advertising keywords; Synonym equivalence matching: Synonym clustering is achieved by calculating the Euclidean distance between words in the semantic space; Feature Engineering : Constructing a vector space through feature engineering; XGBoost predicts probabilities: The XGBoost gradient boosting decision tree model is used to iteratively build a decision tree by fitting the residuals of the previous trees. The objective function is to minimize the residuals containing the first-order gradient. Second gradient and regularization term loss function Controlling model complexity; Classification decision: Intent labels are assigned based on predicted probabilities. When the probability is greater than 0.7, it is judged as an immediate purchase; between 0.4 and 0.7 is the decision period; and less than or equal to 0.4 is the interest exploration period.

3. The advertising design and display system based on big data according to claim 1, characterized in that, The specific steps taken by the scene perception layer to identify the user's scene are as follows: Temporal feature engineering: Constructing a multi-dimensional feature vector containing normalized velocity values, network encoding, time periodic features, and normalized temperature and humidity values; Employing a sliding window strategy to process non-uniform length temporal data and capture contextual dependencies. LSTM-CRF model: A hybrid model based on LSTM gating mechanism and CRF sequence optimization for dynamic scene recognition, where LSTM handles long-term dependencies and CRF corrects local decision contradictions; Environmental data optimization: Environmental sensor data is introduced and the initial scene probability is fused with environmental evidence through a Bayesian update framework, and a dynamic weight adjustment strategy is adopted to optimize the final scene label accuracy.

4. The advertising design and display system based on big data according to claim 1, characterized in that, The specific steps of the feature recombination layer in performing high-order feature interaction through feature engineering and deep learning are as follows: Cartesian product explicit combination: Using the mathematical Cartesian product operation, the user profile, spatiotemporal context, and behavioral sequence feature domains are fully cross-combined to generate explicit combination features of order three or higher; DeepFM Implicit Cross Modeling: Through a joint architecture of factorization machine and deep neural network, the factorization machine part uses vector dot product to calculate second-order feature crosses to capture the linear relationship between features, while the deep neural network part learns higher-order feature interactions through multi-layer nonlinear transformations to discover deep, nonlinear implicit association patterns and generate implicit cross features with business interpretability.

5. The advertising design and display system based on big data according to claim 1, characterized in that, The specific steps for filtering ads using the deep matching layer are as follows: Feature Input and Forward Propagation in Deep Neural Networks: Explicit Feature Vectors With implicit feature vectors Concatenate into a joint feature vector The input-degree neural network model consists of L fully connected layers. The first L−1 layers use the ReLU activation function, and the last layer outputs the predicted values ​​of the three nodes. The output is constrained to the [0,1] interval by the Sigmoid function. The forward propagation process is represented as follows: ,in, For the Sigmoid function, This is the output of the l-th layer. Let be the weight matrix of the l-th layer. Let l be the bias vector of the l-th layer. This is the predicted value for semantic matching degree. The predicted value is the cross-feature score. The predicted value for user intent weight; Multi-task learning: Simultaneously optimize three objectives—semantic matching degree, cross-feature score, and user intent weight—through multi-task learning, and generate a comprehensive matching score through a weighted fusion strategy; The selection strategy based on comprehensive scores is as follows: the scores are compressed to the 0-1 range using the Sigmoid function, and the ads are selected according to the preset threshold. Scores greater than 0.8 are directly entered into the candidate pool, scores between 0.5 and 0.8 need to be sorted again based on scene adaptability, and scores below 0.5 are eliminated.

6. The advertising design and display system based on big data according to claim 1, characterized in that, The specific steps involved in quantifying the contribution of different ad touchpoints to the final conversion in the value measurement layer are as follows: Weighting Calculation: Let the ad type be T∈{Brand Ads, Performance Ads}, and the overall value score be calculated. for ,in, For the growth rate of advertising search volume, For social media buzz, LTV is user lifetime value, CTR is click-through rate, and CVR is conversion rate; Secondary ranking strategy: When the comprehensive value score When ∈[0.5,0.8], a secondary sort is performed based on scene adaptability: , The scene weight is used for scene adaptation, which is the degree to which the ad content matches the current scene. Shapley Value Attribution Analysis: This analysis quantifies the contribution of different ad touchpoints to the final conversion using Shapley Value. Let the set of ad touchpoints be N = {1, 2, ..., n}, and the Shapley Value of touchpoint i be... for , The marginal contribution of the contact subset S, The number of combinations represents the size of all possible subsets. If the Shapley Value of search ads exceeds 50%, the search ad budget will be increased, and the landing page conversion rate will be optimized. If the Shapley Value of display ads is less than 20%, but the CTR is artificially high, the display ad bid will be reduced, and the creative relevance will be optimized.

7. The advertising design and display system based on big data according to claim 6, characterized in that, The specific steps for ad rendering and delivery in the execution layer are as follows: Max-heap priority queue construction: Ads are constructed into a max-heap based on their overall matching scores. The max-heap satisfies the condition that the overall matching score is greater than or equal to the overall score of each ad. When a new ad enters the queue or the score of an existing ad is updated, the heap structure is maintained through an up-filtering operation. Dynamic frequency control: Let the LTV of user u be... The frequency control parameters for advertisement a are: ,in, For LTV stratification threshold, Based on the upper limit of basic frequency control, To set the maximum frequency control limit, record the number of times ad a is displayed for user u within the last T days, C(u,a). If so, display is allowed; otherwise, skip. The time decay factor, The time since the most recent display; LTV layered display: Greater than or equal to Indicating high value, the frequency control upper limit f(u,a) is: The advertising content preference is personalized recommendations and brand advertising; exist and The value between them is represented by the frequency control upper limit f(u,a) as follows: The advertising content preference is for category-related ads and light promotions. Less than or equal to Indicating low value, the frequency control upper limit f(u,a) is The advertising content preferences are high-discount ads and new user gift packs.

8. A big data-based advertising design and display method, using the big data-based advertising design and display system as described in claim 1, characterized in that, Includes the following steps; Multi-source data acquisition and preprocessing: Integrating multiple data sources, using Kafka real-time stream processing technology to build a high-concurrency data pipeline to buffer and distribute user behavior logs, spatiotemporal environmental data and social media volume, and combining the Spark structured computing engine to complete data cleaning, format standardization and lightweight feature extraction; Deep User Intent Analysis: Based on the BERT-Chinese pre-trained language model, user search terms and advertising copy are transformed into high-dimensional semantic vectors. The semantic matching degree between the two is quantified by cosine similarity calculation. At the same time, XGBoost gradient boosting decision tree is used to classify user behavior features and provide dynamic weight coefficients for the deep matching layer. Dynamic scene perception: By integrating GPS speed, network type and time series data, an LSTM-CRF time series model is constructed to dynamically identify the scene in which the user is located, and scene labels are further optimized by combining environmental sensor data; Intelligent feature recombination: User profiles, spatiotemporal context, and behavioral sequences are combined using Cartesian products to generate explicit features, and then combined with the DeepFM automatic feature cross-model to mine implicit related features; Deep matching and comprehensive score generation: Construct a DNN multi-task learning model, integrate explicit features and implicit association features, and optimize semantic matching degree, cross feature score (DeepFM output) and user intent weight to generate a comprehensive matching score in the range of 0-1. Ads with a score > 0.8 directly enter the candidate pool, and ads with a score of 0.5-0.8 are sorted in a second time based on scene adaptability. Multi-objective value assessment: By collecting real-time feedback metrics through front-end tracking, brand value metrics from search engine data / social media APIs, user quality metrics from order data / CRM systems, and negative feedback metrics from user behavior logs, a multi-objective evaluation indicator system is constructed. Combined with dynamic weight allocation and Shapley Value attribution analysis, the contribution of different advertising touchpoints to the final conversion is quantified. Ad design and display: A max-heap priority queue is constructed based on the comprehensive matching score to ensure that high-value ads are displayed first. Dynamic frequency control is implemented in combination with the user LTV stratification strategy, and dynamic ad templates are generated according to the scenario adaptation strategy.

Images