Big data-based advertisement putting intelligent bidding method and system

Abstract

The invention provides an advertisement putting intelligent bidding method and system based on big data. The method comprises the following steps: S1, carrying out multi-source data collection and preprocessing; s2, dynamically constructing a user portrait; s3, carrying out the deep analysis of the context environment; s4, a step of carrying out competition situation multi-dimensional evaluation; s5, training and optimizing the prediction model; s6, making a dynamic bidding strategy; s7, performing real-time bidding execution and adaptive adjustment; s8, carrying out performance monitoring and multi-dimensional feedback collection; s9, carrying out model updating and incremental learning; and S10, carrying out anomaly detection and self-healing processing. The advertisement putting intelligent bidding method based on big data has the advantages that the advertisement putting accuracy is improved, CTR and CVR are predicted to be improved by 15-20% through multi-dimensional data fusion and dynamic user portraits, and irrelevant advertisement putting is reduced.

Description

Technical Field

[0001] This invention specifically relates to an intelligent bidding method and system for advertising placement based on big data. Background Technology

[0002] In the digital advertising field, intelligent bidding is a core component. It automatically adjusts ad bids through big data and machine learning technologies to maximize advertisers' return on investment (ROI). Currently, advertising platforms (such as Google Ads and Facebook Ads) generally use bidding models based on historical data, such as using click-through rate (CTR) or conversion rate (CVR) predictions to set bids. These methods rely on simple regression models or rule engines and often fail to fully integrate multi-source data (such as user behavior, contextual environment, and competitive dynamics), resulting in inefficient bidding. The main challenges of existing technologies include: First, severe data silos, where user profiles, environmental factors, and competitive information are processed in isolation, lacking unified integration and leading to biased bidding decisions; second, insufficient real-time performance, with most models relying on batch updates, failing to adapt to rapidly changing advertising auction environments, resulting in wasted budgets or lost opportunities; third, over-reliance on static thresholds and a lack of dynamic adjustment mechanisms, causing bidding strategies to fail when market conditions change abruptly (e.g., breaking news events); fourth, coarse competitive landscape assessment, considering only direct bidders and ignoring indirect factors (e.g., brand influence or seasonal trends), leading to overbidding or underbidding; and finally, weak model generalization ability and poor prediction accuracy for new users or long-tail scenarios. These problems collectively result in decreased advertiser ROI, platform revenue losses, and impaired user experience (e.g., the proliferation of irrelevant ads). Therefore, the industry urgently needs an intelligent bidding method that can deeply integrate multi-dimensional data, achieve real-time dynamic optimization, and possess strong adaptability. Summary of the Invention

[0003] The purpose of this invention is to address the shortcomings of existing technologies by providing a big data-based intelligent bidding method for advertising, which can effectively solve the aforementioned problems.

[0004] To achieve the above requirements, the technical solution adopted by the present invention is: to provide a big data-based intelligent bidding method for advertising, which includes the following steps:

[0005] S1: Steps for multi-source data collection and preprocessing;

[0006] S2: Steps for dynamically building user profiles;

[0007] S3: Steps for performing in-depth contextual analysis;

[0008] S4: Steps for conducting a multi-dimensional assessment of the competitive landscape;

[0009] S5: Steps for training and optimizing the prediction model;

[0010] S6: Steps for developing a dynamic bidding strategy;

[0011] S7: Steps for real-time bidding execution and adaptive adjustment;

[0012] S8: Steps for performance monitoring and multi-dimensional feedback collection;

[0013] S9: Steps for model updating and incremental learning;

[0014] S10: Steps for anomaly detection and self-healing.

[0015] The advantages of this big data-based intelligent bidding method for ad placement are as follows:

[0016] 1. Improve the accuracy of ad targeting. Through multi-dimensional data fusion and dynamic user profiling, CTR and CVR are expected to increase by 15-20%, while reducing irrelevant ad placements.

[0017] 2. Enhanced real-time adaptability: Through streaming processing and incremental learning, bidding latency is reduced to milliseconds, adapting to sudden market changes and reducing budget waste by 10-15%.

[0018] 3. Optimize risk control by using complex statistical models and risk perception strategies to reduce bid volatility and improve ROI stability.

[0019] 4. Improve system robustness: Through anomaly detection and self-healing, system downtime is reduced and reliability is enhanced.

[0020] 5. Reduce operating costs; automated closed-loop optimization reduces manual intervention and improves scalability. Attached Figure Description

[0021] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, use the same reference numerals to denote the same or similar parts. The illustrative embodiments of this application and their descriptions are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0022] Figure 1 A schematic flowchart of a big data-based intelligent bidding method for ad delivery according to an embodiment of this application is shown. Detailed Implementation

[0023] To make the objectives, technical solutions and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and specific embodiments.

[0024] In the following description, references to "an embodiment," "an embodiment," "an example," "example," etc., indicate that the described embodiment or example may include a particular feature, structure, characteristic, property, element, or limitation, but not every embodiment or example necessarily includes that particular feature, structure, characteristic, property, element, or limitation. Furthermore, the repeated use of the phrase "an embodiment according to this application," while possibly referring to the same embodiment, does not necessarily refer to the same embodiment.

[0025] For simplicity, certain technical features known to those skilled in the art are omitted in the following description.

[0026] According to one embodiment of this application, a smart bidding method for advertising based on big data is provided, such as... Figure 1 As shown, it includes the following steps:

[0027] Step S1: Multi-Source Data Collection and Preprocessing: This step aims to collect raw data from heterogeneous sources and perform cleaning, normalization, and integration to provide high-quality input for subsequent analysis. Data sources include: user behavior data (such as clickstreams, browsing history, and social interactions), contextual data (such as web page content, time, and geographic location), competitive data (such as competitor bids and ad inventory), and external data (such as economic indicators and weather events). Operationally, data is first collected in real time through API interfaces and streaming platforms (such as Apache Kafka) to ensure low latency. For example, user behavior data is extracted from website logs, contextual data is obtained through web crawlers, and competitive data is subscribed from ad exchange platforms. Preprocessing includes: data cleaning (removing duplicates and handling outliers), data transformation (converting unstructured text into vectors using NLP techniques), data normalization (scaling features at different scales to a uniform range, such as Min-Max normalization), and data integration (associating multi-source data through entity resolution, such as linking behavior and environmental data using user IDs). Working Principle: This step is based on a big data processing framework (such as Hadoop or Spark), using distributed computing to process massive amounts of data, ensuring high throughput and fault tolerance. The core of data preprocessing is reducing noise and improving consistency. For example, outliers are identified and addressed using anomaly detection algorithms (such as Isolation Forest) to avoid bias in subsequent models. Furthermore, data integration uses graph databases (such as Neo4j) to build a relational network, revealing implicit connections between users, environment, and competition. The output of this step is a clean, structured, multi-dimensional dataset, which serves as the input for step 2. Its innovation lies in dynamic data prioritization, i.e., adjusting processing resources in real time based on data freshness and importance. For example, real-time data is processed first, while historical data is processed in batches, solving the data latency problem of existing methods.

[0028] Step S2: Dynamic User Profile Construction: This step builds dynamically updated user profiles based on the preprocessed data from Step 1 to accurately depict user interests, intentions, and value. Operationally, user features are first extracted from the dataset, including demographic (e.g., age, gender), behavioral (e.g., click frequency, dwell time), psychographic (e.g., interest tags, sentiment tendencies), and transactional (e.g., purchase history) dimensions. Then, clustering algorithms (e.g., K-means) are used to group users, for example, into "high-value converters" or "browsing users." Dynamic construction is achieved through a real-time update mechanism: as new data flows in, user profiles are adjusted through incremental learning (e.g., online clustering) rather than batch reconstruction. For example, using a streaming K-means algorithm, cluster centers are updated each time new user behavior is received. How it works: The core of user profiling is feature engineering and similarity calculation. Feature engineering uses embedding techniques (e.g., Word2Vecfor text data) to reduce the dimensionality of high-dimensional features, improving computational efficiency. Similarity calculation uses cosine similarity or the Jaccard index to ensure the accuracy of user grouping. Furthermore, this step introduces a time decay factor, giving higher weight to recent behaviors to capture user interest drift. The output is a dynamic profile vector for each user, including interest score, conversion probability, and lifetime value (LTV) estimate. The innovation of this step lies in multimodal fusion, which integrates text, image, and sequence data (such as video browsing history) and extracts cross-modal features through deep learning models (such as Transformer), addressing the problem of single-profile profiling in existing methods. This output serves as input to step 3, ensuring that the contextual analysis is well-founded.

[0029] Step S3: In-depth Contextual Analysis: This step utilizes the user profile from Step 2, combined with the contextual data from Step 1, to analyze the semantic and emotional impact of the ad display environment to optimize bidding strategies. Operationally, it first parses webpage content, device type, timestamps, and location information. For example, NLP techniques (such as the BERT model) are used to model the themes of the webpage text and extract keywords (such as "technology news" or "shopping guide"); simultaneously, computer vision analysis is used to analyze page images and identify visual context (such as the prominence of ad placements). Sentiment analysis is used to assess environmental sentiment (positive, negative, or neutral), for example, using an LSTM network to process user review data. How it works: This step is based on context-aware computation, quantifying environmental factors into feature vectors. For example, an attention mechanism is used to focus on key contextual elements to calculate the matching degree between the environment and the user profile. The matching degree formula is:

[0030] ;

[0031] in Indicates user characteristics, Representing contextual features, is the weight, and sim is the similarity function (e.g., Euclidean distance). The output is the environment score, representing the attractiveness of the environment to the target user. The innovation of this step lies in real-time context adaptation, that is, when the environment changes abruptly (e.g., breaking news), the score is dynamically adjusted through an event detection algorithm (e.g., CUSUM) to avoid bidding in an unsuitable environment. This output serves as the input to step 4, providing an environmental benchmark for competition evaluation.

[0032] Step S4: Multi-dimensional Assessment of Competitive Landscape: This step, based on the environmental score in Step 3, assesses the intensity of competition and multi-dimensional influencing factors in the advertising auction. Operationally, it first extracts competitor bids, ad creatives, and market share information from the competitive data in Step 1. Then, it uses game theory models to simulate competitive dynamics; for example, it constructs a non-cooperative game framework, treating each advertiser as a player and bids as strategies. Multi-dimensional assessment includes direct competition (competitor bid distribution) and indirect competition (such as brand influence and seasonal trends). For example, it predicts competitive fluctuations through time series analysis and estimates future bid trends using an ARIMA model. How it works: The core of this step is the calculation of a competition index, which integrates bid levels, ad quality, and external factors. Specifically, it uses principal component analysis (PCA) to reduce the dimensionality of competitive features and then estimates competitive intensity through a regression model. The output is a dynamic competition index representing the intensity of the current auction. The innovation of this step lies in introducing network effects, i.e., using graph theory to analyze the connections between competitors (such as shared target audiences) to calculate the impact of competitive propagation, solving the problem of isolated assessment in existing methods. This output serves as input to Step 5, ensuring that the predictive model considers competitive constraints.

[0033] Step S5: Predictive Model Training and Optimization: This step, based on the outputs of the previous four steps (user profile, environment score, competition index), trains a multidimensional predictive model to estimate ad click-through rate (CTR) and conversion rate (CVR), providing a foundation for bidding decisions. Operationally, a training dataset is first constructed, containing historical bidding results, user behavior, and environment characteristics. Then, a complex statistical model—the Bayesian Hierarchical Model (BHM)—is used, which can handle data heterogeneity and uncertainty. The BHM structure includes multiple levels: the first level is individual-level (user-environment pairs), the second level is group-level (user clusters), and the third level is population-level (global trend). Model training uses Markov Chain Monte Carlo (MCMC) method for parameter estimation to ensure the convergence of the posterior distribution. Working principle: The BHM formula is as follows:

[0034] ;

[0035] ;

[0036] The meaning of each symbol is as follows:

[0037] : A binary response variable representing a click or conversion event of user i in environment j (1 indicates that it occurred, 0 indicates that it did not occur).

[0038] : The probability of user i clicking or converting in environment j.

[0039] : Global intercept term, representing the baseline probability.

[0040] , Fixed effects coefficients, representing user feature vectors respectively. (e.g., interest scores) and environmental feature vectors The impact of (such as environmental ratings) on probability.

[0041] Random effects represent specific shifts in environment j that follow a normal distribution. ~N(0, ), to capture variations between environments.

[0042] Random effects represent a specific offset of user i, which follows a normal distribution. ~N(0, ), to capture variations among users.

[0043] The residual term follows a normal distribution. ~N(0, ), representing unobserved noise. During training, parameters are updated using Gibbs sampling, and the posterior distribution is approximated through iterative sampling. Model optimization selects hyperparameters (such as prior distributions) through cross-validation and evaluates goodness of fit using information criteria (such as WAIC). The output is the predicted probability distribution, including point estimates and uncertainty intervals. The innovation of this step lies in introducing multi-level random effects, effectively handling data sparsity issues (such as new users) and improving prediction robustness. This output serves as the input to step 6, providing the probabilistic basis for the dynamic policy.

[0044] Step S6: Dynamic Bidding Strategy Formulation: This step, based on the predicted probabilities from Step 5, formulates a real-time bidding strategy to balance ROI and risk. Operationally, first, a target function is defined, such as maximizing expected value (e.g., eCPM = CTR * CVR * Bid). Then, reinforcement learning (e.g., Q-learning) is used to dynamically adjust the bid: the bidding environment is viewed as a Markov Decision Process (MDP), where the states include user profiles, environment scores, and competition indices; the action is the bid amount; and the reward is the ad performance (clicks or conversions). Strategy formulation involves iteratively updating the Q-table, with the Q-value update formula as follows:

[0045] ;

[0046] Where s represents the state, a represents the action, and r represents the immediate reward. γ is the learning rate, and γ is the discount factor. How it works: The core of this step is multi-objective optimization, simultaneously considering budget constraints, competition limitations, and user value. For example, constraints are incorporated into the objective function using Lagrange relaxation to solve for the optimal bid. The output is a policy function that maps the state to the bidding decision. The innovation of this step lies in introducing a risk perception mechanism, using Conditional Value at Risk (CVaR) to quantify losses under uncertainty and avoid overbidding. This output serves as the input to step 7, ensuring that bidding execution is based on evidence.

[0047] Step S7: Real-time Bidding Execution and Adaptive Adjustment: This step, based on the strategy in Step 6, executes bids in the advertising auction and fine-tunes them based on real-time feedback. Operationally, the strategy function is first deployed to the bidding engine (such as an RTB system). When an auction request arrives, the engine queries the current state (user, environment, competition), calculates the optimal bid, and submits it. Adaptive adjustment is achieved through online learning: for example, using Thompson sampling to sample probabilities from a predicted distribution and dynamically adjust the bid to explore and leverage trade-offs. How it works: This step is based on real-time stream processing, ensuring low-latency response (milliseconds). For example, Apache Flink is used to process the auction stream, updating the bid every millisecond. The output is the actual bid record and auction result. The innovation of this step lies in introducing a context bandit algorithm, using the environment score from Step 3 as context to improve bidding accuracy. This output serves as input to Step 8, providing data for monitoring.

[0048] Step S8: Performance Monitoring and Multi-Dimensional Feedback Collection: This step, based on the bidding results from Step 7, monitors ad performance and collects feedback data. Operationally, monitoring metrics are first defined, such as CTR, CVR, ROI, and wasted budget percentage. Then, these metrics are visualized in real-time using dashboards (such as Grafana), and alert thresholds are set (e.g., triggering when ROI falls below a threshold). Feedback collection includes user feedback (e.g., survey ratings) and system feedback (e.g., auction win / loss records). How it works: This step uses a time-series database (such as InfluxDB) to store monitoring data and identifies performance bottlenecks through correlation analysis. For example, Pearson correlation coefficients are used to analyze the relationship between bids and performance. The output is a performance report and a feedback dataset. The innovation of this step lies in introducing multi-source feedback fusion, integrating explicit (user ratings) and implicit (behavioral data) feedback to address the problem of single feedback in existing methods. This output serves as input to Step 9, providing a basis for model updates.

[0049] Step S9: Model Update and Incremental Learning: This step updates the prediction model from Step 5 based on the feedback data from Step 8, ensuring adaptation to the dynamic environment. Operationally, incremental learning algorithms (such as online gradient descent) are first used to update the model parameters, avoiding full retraining. For example, for the BHM model, variational inference is used to approximate the posterior, updating the distribution with each new data received. Working principle: This step is based on a continuous learning framework to prevent catastrophic forgetting. For example, elastic weight merging (EWC) is used to protect important parameters. The output is the updated model version. The innovation of this step lies in automatic model selection, comparing different models (such as BHM vs. neural networks) through A / B testing and dynamically switching to the optimal model. This output serves as input to Step 10, ensuring continuous system optimization.

[0050] Step S10: Anomaly Detection and Self-Healing: This step, based on the updated model from Step 9, detects and handles system anomalies, such as sudden bid spikes or data contamination. Operationally, anomaly detection algorithms (such as Isolation Forest or LSTM-AE) are first used to identify anomalous patterns; for example, a sudden bid surge may indicate fraud. Then, self-healing is performed by automatically adjusting parameters through a rule engine or reinforcement learning, such as temporarily lowering bids or switching to a backup model. How it works: This step is based on the concept of an autonomous system, achieving minimal human intervention. The output is a system status report and a record of corrective actions. The innovation of this step lies in multimodal anomaly detection, integrating numerical, sequential, and graph data to improve detection accuracy. The method concludes with this step, forming a closed-loop optimization.

[0051] According to one embodiment of this application, the big data-based intelligent bidding method for advertising addresses several key technical issues in the field of digital advertising bidding. First, it breaks down data silos through multi-source data fusion in step 1 and dynamic user profiling in step 2, achieving unified processing of user, environmental, and competitive data, thus solving the problem of inaccurate bidding caused by incomplete data in existing methods. For example, traditional methods rely solely on historical click data, ignoring contextual changes, while this method, through deep environmental analysis in step 3, dynamically assesses the semantic impact of ad display scenarios, avoiding bidding in unsuitable environments and improving ad relevance. Second, this method introduces uncertainty and risk perception through a complex statistical model (Bayesian hierarchical model) in step 5 and a dynamic strategy in step 6, solving the bidding fluctuation problem caused by excessive reliance on point estimation in existing models. Traditional methods such as logistic regression cannot handle data sparsity, while BHM captures individual and group variations through random effects, improving the prediction accuracy for new users or long-tail scenarios. Third, this method achieves millisecond-level dynamic adjustment through real-time bidding adaptation in step 7 and incremental learning in step 9, solving the latency problem caused by batch updates in existing systems. For example, in the event of an emergency, traditional models may only be updated every few hours, while this method adapts in real time through stream processing, reducing budget waste. Fourth, this method comprehensively handles direct and indirect competitive factors through multi-dimensional competition evaluation in step 4 and anomaly detection in step 10, and automatically corrects anomalies, solving the problems of coarse competition evaluation and system fragility in existing methods. Overall, this method improves bidding efficiency, ROI, and system robustness, bringing significant value to advertisers and platforms.

[0052] The embodiments described above are merely examples of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the claims.

Claims

1. A big data-based advertisement delivery intelligent bidding method, characterized in that, The method comprises the following steps: S1: performing multi-source data collection and preprocessing; S2: performing dynamic user portrait construction; S3: performing in-depth analysis of the context environment; S4: performing multi-dimensional evaluation of the competitive situation; S5: performing prediction model training and optimization; S6: performing dynamic bidding strategy formulation; S7: performing real-time bidding execution and adaptive adjustment; S8: performing performance monitoring and multi-dimensional feedback collection; S9: performing model updating and incremental learning; S10: performing anomaly detection and self-healing processing. 2.The big data based advertisement launching intelligent bidding method according to claim 1, characterized in that, The step S1 specifically comprises: Collecting raw data from heterogeneous sources and performing cleaning, normalization and integration to provide high-quality input for subsequent analysis, the data sources include user behavior data, context data, competition data and external data, first, real-time data collection is performed through API interface and streaming processing platform to ensure low latency, preprocessing includes data cleaning, data conversion, data normalization and data integration, based on big data processing framework, massive data is processed through distributed computing to ensure high throughput and fault tolerance, the core of data preprocessing is to reduce noise and improve consistency, data integration uses graph database to build relationship network to reveal the implicit association of user-environment-competition, the output is clean, structured multi-dimensional data set, dynamic data priority allocation, that is, real-time adjustment of processing resources according to data freshness and importance. 3.The big data based advertisement launching intelligent bidding method according to claim 1, characterized in that, The step S2 specifically comprises: Based on the preprocessed data, a dynamically updated user portrait is constructed to accurately depict user interest, intention and value, first, user features are extracted from the data set, including demographic, behavioral, psychographic and transactional dimensions, clustering algorithm is used to group users, and dynamic construction is realized through real-time updating mechanism, when new data flows in, user portrait is adjusted through incremental learning instead of batch reconstruction, stream K-means algorithm is used, and clustering center is updated every time new user behavior is received, the core of user portrait is feature engineering and similarity calculation, feature engineering uses embedding technology to reduce high-dimensional features, and calculation efficiency is improved, similarity calculation uses cosine similarity or Jaccard index to ensure the accuracy of user grouping, time decay factor is introduced to give higher weight to recent behavior to capture user interest drift, the output is a dynamic portrait vector of each user, including interest score, conversion probability and life cycle value estimation, multi-modal fusion, that is, integrating text, image and sequence data, cross-modal features are extracted through deep learning model to solve the single portrait problem of existing methods. 4.The big data based advertisement launching intelligent bidding method according to claim 1, wherein, The step S3 specifically comprises: Using user portrait, combined with context data, analyze the semantic and emotional influence of the advertising display environment to optimize the bidding strategy, first parse the webpage content, device type, timestamp and location information, use NLP technology to model the theme of webpage text and extract keywords; at the same time, analyze the page image through computer vision to identify the visual context, and use emotion analysis to evaluate the environmental emotion, use LSTM network to process user comment data, based on context perception calculation, quantify the environmental factors into feature vector, focus on key context elements through attention mechanism, calculate the matching degree of environment and user portrait, the matching degree formula is: ; wherein denotes a user feature, denotes a context feature, is a weight, and sim is a similarity function, such as the Euclidean distance; The output is the environment score, which represents the attractiveness of the environment to the target user, real-time context adaptation, that is, when the environment changes, the score is dynamically adjusted through event detection algorithm to avoid bidding in an inappropriate environment. 5.The big data based advertisement launching intelligent bidding method according to claim 1, wherein, The step S4 specifically comprises: Based on the environment score, evaluate the competition intensity and multi-dimensional influence factors in the ad auction, first extract the competitor bid, ad creative and market share information from the competition data, use game theory model to simulate the competition dynamics, build a non-cooperative game framework, regard each advertiser as a player, and the bid as a strategy, multi-dimensional evaluation includes direct competition and indirect competition, predict competition fluctuations through time series analysis, use ARIMA model to estimate future bid trend, competition index calculation, this index integrates bid level, ad quality and external factors, use principal component analysis to reduce the dimension of competition characteristics, then estimate the competition intensity through regression model, the output is a dynamic competition index, which represents the intensity of the current auction, introduce network effect, that is, through graph analysis the association between competitors, calculate the competition propagation influence, solve the problem of isolated evaluation in existing methods. 6.The big data based advertisement launching intelligent bidding method according to claim 1, wherein, The step S5 specifically comprises: Based on the output of the first four steps, train a multi-dimensional prediction model to estimate the ad click-through rate and conversion rate, which provides the basis for bidding decision, first build a training data set containing historical bidding results, user behavior and environmental characteristics, use a complex statistical model, Bayesian hierarchical model, which can handle data heterogeneity and uncertainty, the structure of BHM includes multiple levels: the first level is individual level, the second level is group level, and the third level is total level, model training is performed through Markov chain Monte Carlo method to estimate parameters, ensure the convergence of posterior distribution, the formula of BHM is as follows: ； ； Wherein, the meaning of each symbol is: : Binary response variable indicating a click or conversion event for user i in environment j, 1 indicates it occurred, 0 indicates it did not; : probability of click or conversion by user i in environment j; : global intercept term representing a baseline probability; , : fixed effect coefficients representing the user feature vector , interest score and environment feature vector , the effect of the environment score on the probability; Random effects represent specific shifts in environment j that follow a normal distribution. ~N(0, ), to capture variations between environments; Random effects represent a specific offset of user i, which follows a normal distribution. ~N(0, ), capturing variations among users; The residual term follows a normal distribution. ~N(0, ), indicating unobserved noise; During the training process, the parameters are updated using Gibbs sampling, the posterior distribution is approximated by iterative sampling, the model is optimized by cross-validation to select hyperparameters, and the goodness of fit is evaluated using information criterion, the output is the prediction probability distribution, including point estimate and uncertainty interval, introduce multi-level random effect, effectively handle data sparsity problem, improve prediction robustness. 7.The big data based advertisement launching intelligent bidding method according to claim 1, wherein, The step S6 specifically comprises: Based on the predicted probability, a real-time bidding strategy is formulated to balance ROI and risk. First, define the objective function, and use reinforcement learning to dynamically adjust the bid. The bidding environment is considered as a Markov decision process. The state includes user profile, environment score, and competition index. The action is the bid amount, and the reward is the ad effect. The strategy is updated by value iteration to update the Q table. The Q value update formula is: ； where s is state, a is action, r is immediate reward, is learning rate, γ is discount factor, multi-objective optimization, considering budget constraint, competition restriction and user value, through Lagrange relaxation method, the constraints are integrated into the objective function, the optimal bid is solved, the output is a strategy function, which maps the state to the bid decision, the risk perception mechanism is introduced, the conditional value at risk is used to quantify the loss under uncertainty, and over-bidding is avoided. 8.The big data based advertisement launching intelligent bidding method according to claim 1, wherein, The step S7 specifically includes: Based on the strategy, execute the bid in the ad auction, and fine-tune according to real-time feedback. First, deploy the strategy function to the bidding engine. When the auction request arrives, the engine queries the current state, calculates the optimal bid, and submits it. Adaptive adjustment is achieved through online learning. Use Thompson sampling to sample probabilities from the predicted distribution, dynamically adjust the bid to explore-exploit trade-off, ensure low-latency response based on real-time stream processing, use Apache Flink to process the auction stream, update the bid every millisecond, the output is the actual bid record and the auction result, introduce the context bandit algorithm, use the environment score as the context, and improve the accuracy of the bid. 9.The big data based advertisement launching intelligent bidding method according to claim 1, wherein, The step S8 specifically includes: Based on the bidding result, monitor the ad effect and collect feedback data. First, define the monitoring indicators, CTR, CVR, ROI, and waste budget ratio. Visualize the indicators in real time through the dashboard and set alarm thresholds. Feedback collection includes user feedback and system feedback. Use a time series database to store monitoring data and identify performance bottlenecks through correlation analysis. Use the Pearson correlation coefficient to analyze the relationship between bidding and effect. The output is the performance report and feedback dataset. Introduce multi-source feedback fusion, which integrates explicit and implicit feedback to solve the single feedback problem of existing methods. The step S9 specifically includes: Based on the feedback data, update the prediction model to ensure adaptation to the dynamic environment. First, use incremental learning algorithms to update model parameters to avoid full retraining. For BHM model, approximate the posterior through variational inference, update the distribution every time new data is received. Based on the continuous learning framework, prevent catastrophic forgetting. Use elastic weight consolidation to protect important parameters. The output is the updated model version. Automatic model selection compares different models through A / B testing and dynamically switches the optimal model. The step S10 specifically includes: Based on the updated model, detect and handle system anomalies such as bid spikes or data pollution. First, use anomaly detection algorithms to identify abnormal patterns. A sudden surge in bidding may indicate fraud. Self-healing processing automatically adjusts parameters through a rule engine or reinforcement learning, temporarily reduces bidding or switches to a backup model. Based on the autonomous system concept, achieve minimal human intervention. The output is the system status report and the record of corrective measures. Multi-modal anomaly detection integrates numerical, sequential, and graph data to improve detection accuracy. 10.A big data-based advertisement delivery intelligent bidding system, characterized in that: Use the big data-based ad bidding intelligent bidding method as claimed in claims 1-9.

Images