Intelligent advertisement putting system based on big data analysis

Abstract

The invention discloses an intelligent advertisement putting system based on big data analysis, and the system comprises the following modules: a multi-source data collection module which is used for collecting user side behavior data, advertiser demand data and media environment data in real time; the user portrait construction module is used for generating dynamically updated user feature tags based on the multi-source data; through multi-source real-time data acquisition and dynamic user portrait construction, the comprehensiveness and timeliness of user feature recognition are remarkably improved; based on an intelligent matching mechanism of deep learning and a double-tower neural network, high-precision association between advertisements and users is realized; the advertisement putting efficiency and the user experience are optimized by combining a dynamic bidding strategy, frequency control and cross-platform coordination; and real-time iteration and multi-target balance of the strategy are realized by relying on a closed-loop evaluation optimization framework and deep reinforcement learning, and finally, collaboration of ROI (Region of Interest) improvement of advertisers, platform income increase and user advertisement fatigue reduction is achieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of advertising placement, and particularly relates to an intelligent advertising placement system based on big data analysis. Background Art

[0002] Traditional advertising placement systems generally have problems such as single data collection dimension, lagging user portrait update, advertisement matching relying on manual experience, lack of dynamic adjustment ability for bidding strategies, and imperfect effect evaluation mechanisms. As a result, the accuracy of advertising placement is insufficient, the user experience is damaged, the return on investment of advertisers fluctuates greatly, and it is difficult to meet the complex requirements in the multi-platform real-time bidding environment. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to overcome the existing defects and provide an intelligent advertising placement system based on big data analysis, so as to solve the problems of traditional advertising placement systems generally having single data collection dimension, lagging user portrait update, advertisement matching relying on manual experience, lack of dynamic adjustment ability for bidding strategies, and imperfect effect evaluation mechanisms, resulting in insufficient accuracy of advertising placement, damaged user experience, large fluctuations in the return on investment of advertisers, and difficulty in adapting to the complex requirements in the multi-platform real-time bidding environment.

[0004] To achieve the above object, the present invention provides the following technical solution: An intelligent advertising placement system based on big data analysis, comprising the following modules:

[0005] A multi-source data collection module, used for collecting user-side behavior data, advertiser demand data, and media environment data in real time;

[0006] A user portrait construction module, generating dynamically updated user feature tags based on the multi-source data;

[0007] An advertisement intelligent matching module, calculating the matching degree between advertisements and users by using a deep learning model;

[0008] A real-time bidding decision module, executing cross-platform dynamic bidding strategies and frequency control;

[0009] An effect evaluation and optimization module, realizing the closed-loop iterative optimization of the placement strategy.

[0010] Preferably, the multi-source data collection module includes:

[0011] A user behavior collection sub-module, collecting the behavior sequence data of users on mobile terminals through an embedded SDK, including click events, page stay duration, and shopping cart operations;

[0012] The advertiser demand analysis sub-module is used to analyze the structured demand data and unstructured creative materials provided by advertisers, including budget constraints, target population characteristics, advertising copy, and multimedia content;

[0013] The media environment perception sub-module integrates natural language processing technology to extract the context semantic features of web pages / videos, including page keywords, video frame features, ad slot size, and location information;

[0014] The data cleaning sub-module adopts a hybrid cleaning mechanism based on a rule engine and an anomaly detection algorithm to denoise, complete, and standardize the collected data.

[0015] Preferably, the user behavior collection sub-module includes:

[0016] The real-time streaming collection unit is a streaming computing engine built based on Apache Flink, supporting the processing of 100,000-level events per second;

[0017] The behavior correlation analysis unit unifies the cross-device behavior data through the user ID mapping table;

[0018] The privacy protection unit uses differential privacy technology to desensitize sensitive data to ensure compliance with GDPR regulations.

[0019] Preferably, the user portrait construction module includes:

[0020] The basic attribute extraction sub-module extracts static features such as gender, age, and region from the user registration information;

[0021] The behavior feature modeling sub-module uses time series analysis algorithms to construct user behavior patterns, including:

[0022] The short-term interest model calculates the real-time interest weight based on the behavior data in the past 7 days;

[0023] The long-term preference model uses the LSTM neural network to mine the behavior patterns over 30 days;

[0024] The consumption ability evaluation sub-module constructs a consumption power index by combining historical order amounts, payment methods, and coupon usage frequencies;

[0025] The scenario perception sub-module identifies the scenario characteristics where the user is located through geographical location data and device sensors.

[0026] Preferably, the behavior feature modeling sub-module adopts a dynamic decay algorithm:

[0027] When the user has no relevant behavior for 24 consecutive hours, the priority of this label is automatically reduced.

[0028] Preferably, the advertising intelligent matching module includes:

[0029] The feature engineering sub-module maps user features and ad features to the same vector space, where:

[0030] The dimension of the user feature vector ≥ 256;

[0031] The dimension of the ad feature vector ≥ 128;

[0032] The deep matching sub-module adopts a two-tower neural network architecture, including:

[0033] The user feature tower is composed of a 3-layer fully connected network, and the activation function is LeakyReLU;

[0034] The ad feature tower integrates CNN for processing image ads and BERT for processing text ads;

[0035] The multi-objective optimization sub-module balances the Pareto front solution sets of the advertiser's ROI, user experience, and platform revenue.

[0036] Preferably, the training process of the deep matching sub-module includes:

[0037] Construct a training set containing 100 million user-ad interaction records;

[0038] Adopt a contrast loss function, and the negative sample sampling ratio is 1:5;

[0039] Update the model parameters through the dynamic weight averaging algorithm, and the learning rate is set to 0.001;

[0040] Perform an online model hot update every 6 hours.

[0041] Preferably, the real-time bidding decision module includes:

[0042] The dynamic bidding sub-module adjusts the bid according to the user value coefficient and ad slot quality;

[0043] The frequency control sub-module uses a sliding window algorithm to limit the number of user exposures, and the window period is 24 hours;

[0044] The cross-platform coordination sub-module synchronously processes the bidding requests of multiple ad trading platforms (DSPs), and the priority strategies include:

[0045] High-value users are given priority to display high-price ads;

[0046] New ad materials are given priority to obtain test traffic;

[0047] Allocate budgets dynamically according to the platform's historical performance.

[0048] Preferably, the frequency control sub-module realizes:

[0049] Build a user - ad exposure matrix to record the exposure times of each user to various types of ads;

[0050] When the fatigue level exceeds 0.8, suspend the delivery of similar ads for 12 hours.

[0051] Preferably, the effect evaluation and optimization module includes:

[0052] A multi - dimensional analysis sub - module that calculates 12 core indicators such as ad conversion rate, earnings per thousand impressions (eCPM), and user retention rate;

[0053] An anomaly detection sub - module that uses the Isolation Forest algorithm to identify abnormal delivery events, with a detection response time < 500ms;

[0054] A policy iteration sub - module that dynamically adjusts the parameters of an intelligent ad delivery system based on big data analysis based on a deep reinforcement learning framework, including:

[0055] State space: contains more than 50 real - time delivery features;

[0056] Action space: adjusts the matching threshold, bid coefficient, and frequency limit parameters;

[0057] An A / B test sub - module that runs multiple groups of policies in parallel and compares the effect differences, with a decision - making cycle of 6 hours.

[0058] Compared with the prior art, the present invention provides an intelligent ad delivery system based on big data analysis, having the following beneficial effects:

[0059] Through multi - source real - time data collection and dynamic user portrait construction, the present invention significantly improves the comprehensiveness and timeliness of user feature recognition; based on an intelligent matching mechanism of deep learning and dual - tower neural network, it realizes a high - precision association between ads and users; combined with dynamic bidding strategies, frequency control, and cross - platform coordination, it optimizes ad delivery efficiency and user experience; relying on a closed - loop evaluation and optimization framework and deep reinforcement learning, it realizes real - time iteration of strategies and multi - objective balance, and finally achieves the synergy of improving the ROI of advertisers, increasing platform revenue, and reducing user ad fatigue. Detailed implementation manners

[0060] The technical solutions in the embodiments of the present invention are described clearly and completely below. Apparently, the described embodiments are only a part of the embodiments of the present invention, rather than all embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0061] The present invention provides a technical solution: an intelligent ad delivery system based on big data analysis, including the following modules:

[0062] The multi-source data collection module 100 is used to collect user-side behavior data, advertiser demand data, and media environment data in real time;

[0063] The user profile construction module 200 generates dynamically updated user feature tags based on the multi-source data;

[0064] The advertisement intelligent matching module 300 calculates the matching degree between advertisements and users using a deep learning model;

[0065] The real-time bidding decision-making module 400 executes cross-platform dynamic bidding strategies and frequency control;

[0066] The effect evaluation and optimization module 500 realizes the closed-loop iterative optimization of the placement strategy.

[0067] In the present invention, preferably, the multi-source data collection module 100 includes:

[0068] The user behavior collection sub-module 110 collects the behavior sequence data of users on mobile terminals through an embedded SDK, including click events, page stay duration, and shopping cart operations;

[0069] The advertiser demand parsing sub-module 120 is used to parse the structured demand data and unstructured creative materials provided by advertisers, including budget constraints, target population characteristics, advertisement copy, and multimedia content;

[0070] The media environment perception sub-module 130 integrates natural language processing technology to extract the context semantic features of web pages / videos, including page keywords, video frame features, advertisement space size, and position information;

[0071] The data cleaning sub-module 140 adopts a hybrid cleaning mechanism based on a rule engine and an anomaly detection algorithm to denoise, complete, and standardize the collected data.

[0072] In the present invention, preferably, the user behavior collection sub-module 110 includes:

[0073] The real-time streaming collection unit 111 is a streaming computing engine built based on Apache Flink, supporting the processing of 100,000-level events per second;

[0074] The behavior correlation analysis unit 112 uniformly correlates cross-device behavior data through a user ID mapping table;

[0075] The privacy protection unit 113 uses differential privacy technology to desensitize sensitive data to ensure compliance with GDPR specifications.

[0076] In the present invention, preferably, the user profile construction module 200 includes:

[0077] The basic attribute extraction sub-module 210 extracts static features such as gender, age, and region from the user registration information;

[0078] The behavior feature modeling sub-module 220 constructs the user behavior pattern by using the time series analysis algorithm, including:

[0079] The short-term interest model 221 calculates the real-time interest weight based on the behavior data in the recent 7 days;

[0080] The long-term preference model 222 uses the LSTM neural network to mine the behavior patterns exceeding 30 days;

[0081] The consumption ability evaluation sub-module 230 constructs the consumption power index by combining the historical order amount, payment method, and coupon usage frequency;

[0082] The scene perception sub-module 240 identifies the scene features where the user is located through the geographical location data and the device sensor.

[0083] In the present invention, preferably, the behavior feature modeling sub-module 220 adopts the dynamic decay algorithm:

[0084] When the user has no relevant behavior for 24 consecutive hours, the priority of this label is automatically reduced.

[0085] In the present invention, preferably, the advertisement intelligent matching module 300 includes:

[0086] The feature engineering sub-module 310 maps the user features and the advertisement features to the same vector space, where:

[0087] The dimension of the user feature vector ≥ 256;

[0088] The dimension of the advertisement feature vector ≥ 128;

[0089] The deep matching sub-module 320 adopts the dual tower neural network architecture, including:

[0090] The user feature tower 321 is composed of 3 fully connected networks, and the activation function is LeakyReLU;

[0091] The advertisement feature tower 322 integrates CNN to process image advertisements and BERT to process text advertisements;

[0092] The multi-objective optimization sub-module 330 balances the Pareto front solution sets of the advertiser's ROI, user experience, and platform revenue.

[0093] In the present invention, preferably, the training process of the deep matching sub-module 320 includes:

[0094] Construct a training set containing 100 million user-advertisement interaction records;

[0095] The contrast loss function is adopted, and the negative sample sampling ratio is 1:5;

[0096] The model parameters are updated through the dynamic weight averaging algorithm, and the learning rate is set to 0.001;

[0097] The online model is hot-updated every 6 hours.

[0098] In the present invention, preferably, the real-time bidding decision module 400 includes:

[0099] A dynamic bidding sub-module 410 that adjusts the bid according to the user value coefficient and the ad slot quality;

[0100] A frequency control sub-module 420 that uses a sliding window algorithm to limit the user exposure times, and the window period is 24 hours;

[0101] A cross-platform coordination sub-module 430 that synchronously processes the bidding requests of multiple ad trading platforms DSPs, and the priority strategies include:

[0102] High-value users are given priority to place high-price ads;

[0103] New ad materials are given priority to obtain test traffic;

[0104] The budget is dynamically allocated according to the platform historical performance.

[0105] In the present invention, preferably, the frequency control sub-module 420 implements:

[0106] Establish a user-ad exposure matrix to record the exposure times of each user to various types of ads;

[0107] When the fatigue degree exceeds 0.8, the placement of similar ads is paused for 12 hours.

[0108] In the present invention, preferably, the effect evaluation and optimization module 500 includes:

[0109] A multi-dimensional analysis sub-module 510 that calculates 12 core indicators such as the ad conversion rate, earnings per thousand impressions eCPM, and user retention rate;

[0110] An anomaly detection sub-module 520 that uses the isolation forest algorithm to identify abnormal placement events, and the detection response time <500ms;

[0111] A strategy iteration sub-module 530 that dynamically adjusts the parameters of an intelligent advertising placement system based on big data analysis based on a deep reinforcement learning framework, including:

[0112] State space: includes more than 50 real-time placement features;

[0113] Action space: adjust the matching threshold, bid coefficient, and frequency limit parameter;

[0114] The A / B test sub-module 540 runs multiple groups of strategies in parallel and compares the effect differences, with a decision-making cycle of 6 hours.

[0115] Example 1:

[0116] An intelligent advertising delivery system based on big data analysis includes the following implementation processes:

[0117] Step 1: Multi-source data collection: Real-time collection of user behavior data such as clicking on product detail pages, adding to the shopping cart, and favoriting products through an embedded SDK; parsing advertisers' promotion budgets, target audiences (such as women aged 25 - 35), and advertising materials (copywriting + product pictures);

[0118] Step 2: User portrait construction: Basic attribute extraction: gender, age, location; Behavior modeling: Generate a consumption power index based on the browsing record in the past 7 days (short-term interest model) and historical orders (long-term preference model); Scene perception: Identify that the user is currently in the "weekend evening mobile" scenario;

[0119] Step 3: Intelligent advertising matching: Feature engineering: Map user features (256 dimensions) and advertising features (128 dimensions) to a vector space; Dual tower neural network matching: Calculate the matching degree between the user feature tower (fully connected network) and the advertising feature tower (CNN processes product pictures), and preferentially recommend highly relevant promotional advertisements;

[0120] Step 4: Real-time bidding and frequency control: Dynamic bidding: Increase the bid for high-consumption power users by 20%; Frequency limit: The exposure of the same type of advertisement to the same user within 24 hours does not exceed 3 times, and stop delivery after the fatigue degree exceeds the threshold;

[0121] Step 5: Effect optimization: Multi-dimensional analysis: Monitor that the conversion rate increases by 15% and the eCPM increases by 12%; Strategy iteration: Adjust the bid coefficient through deep reinforcement learning, and fully roll out after verification by A / B testing; Effect: The advertiser's ROI increases by 18%, the user click-through rate increases by 25%, and the platform revenue increases by 20%.

[0122] Example 2:

[0123] An intelligent advertising delivery system based on big data analysis includes the following implementation processes:

[0124] Step 1: Media environment perception: NLP extracts video subtitle keywords (such as "fitness tutorials"); CNN analyzes video key frames (identifies fitness equipment scenes); Advertising position information: 15-second pre-roll advertisement, with a full-screen size;

[0125] Step 2. User Portrait Construction: Behavioral Characteristics: The proportion of the viewing duration of fitness videos by the user in the past 30 days is 60%; Scene Perception: The user's frequent gym is located through GPS positioning, and the device sensor detects that it is the "evening exercise period" currently;

[0126] Step 3. Ad Matching and Bidding: The ad feature tower uses BERT to parse fitness brand copywriting to match the user's long-term preferences; Cross-platform Coordination: Prioritize bidding for highly relevant fitness equipment ads among multiple DSPs; Dynamic Bidding: Adjust the bid according to the ad slot quality (full-screen display weight + 1.2);

[0127] Step 4. Frequency and Fatigue Management: The sliding window algorithm limits the exposure of fitness ads to the same user to ≤ 2 times per day; When the fatigue degree > 0.8, replace it with low-intensity nutritional product ads;

[0128] Step 5. Evaluation and Optimization: Anomaly Detection: The Isolation Forest algorithm identifies ad slots with abnormally low conversion rates, and the response time < 300 ms; Policy Iteration: Optimize the test ratio of ad materials based on user retention rate; Effect: The ad conversion rate increases by 22%, the user ad skip rate decreases by 30%, and the platform eCPM increases by 18%.

[0129] Example 3:

[0130] An intelligent advertising delivery system based on big data analysis includes the following implementation processes:

[0131] Step 1. Cross-device Data Association: The user behavior collection sub-module associates multi-device behaviors through the ID mapping table (such as searching for "travel guides" on mobile phones and booking hotels on PCs); The privacy protection unit desensitizes cross-device data to ensure GDPR compliance;

[0132] Step 2. Dynamic Update of User Portrait: Short-term Interest Model: Based on real-time browsing data on the mobile side (the keyword "island tourism" in the past 7 days); Long-term Preference Model: LSTM analyzes the historical orders on the PC side (annual average tourism consumption of 20,000 yuan); Scene Perception: Identify the user's current use of the "mobile outdoor scene" through the device sensor;

[0133] Step 3. Intelligent Matching and Bidding: Feature Engineering: Unify multi-device user features (256 dimensions) to match travel ads (128 dimensions); Multi-objective Optimization: Balance the ROI of travel agencies, user experience (reduce duplicate ads), and platform revenue; Cross-platform Coordination: Synchronously bid for high-end travel product ads for high-value users in the DSP;

[0134] Step 4: Frequency Control and Policy Iteration: The user-advertisement exposure matrix records the exposure times of each device, and the total frequency is limited to 5 times per day; the A / B test sub-module runs the "bid coefficient + 0.1" and "frequency limit - 1" policies in parallel, and selects the optimal solution after 6 hours; Effect: The cross-device advertisement click consistency is increased by 40%, the user fatigue is reduced by 35%, and the advertiser's ROI is increased by 15%.

[0135] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent advertising delivery system based on big data analysis, characterized in that, It includes the following modules: The multi-source data collection module (100) is used to collect user behavior data, advertiser demand data, and media environment data in real time; The user portrait construction module (200) generates dynamically updated user feature tags based on the multi-source data; The advertisement intelligent matching module (300) calculates the matching degree between advertisements and users by using a deep learning model; The real-time bidding decision module (400) executes cross-platform dynamic bidding strategies and frequency control; The effect evaluation and optimization module (500) realizes the closed-loop iterative optimization of the placement strategy.

2. The intelligent advertising delivery system based on big data analysis according to claim 1, wherein The multi-source data collection module (100) includes: The user behavior collection sub-module (110) collects the behavior sequence data of users on mobile terminals through an embedded SDK, including click events, page stay duration, and shopping cart operations; The advertiser demand analysis sub-module (120) is used to analyze the structured demand data and unstructured creative materials provided by advertisers, including budget constraints, target population characteristics, advertisement copywriting, and multimedia content; The media environment perception sub-module (130) integrates natural language processing technology to extract the context semantic features of web pages / videos, including page keywords, video frame features, advertisement space size, and location information; The data cleaning sub-module (140) uses a hybrid cleaning mechanism based on a rule engine and an anomaly detection algorithm to denoise, complete, and standardize the collected data.

3. An intelligent advertising placement system based on big data analysis according to claim 2, characterized in that, The user behavior collection sub-module (110) contains: The real-time streaming collection unit (111), a streaming computing engine built based on Apache Flink, supporting the processing of 100,000-level events per second; The behavior correlation analysis unit (112) uniformly correlates cross-device behavior data through a user ID mapping table; The privacy protection unit (113) uses differential privacy technology to desensitize sensitive data to ensure compliance with GDPR specifications.

4. An intelligent advertising delivery system based on big data analysis according to claim 1, characterized in that The user portrait construction module (200) includes: The basic attribute extraction sub-module (210) extracts static features such as gender, age, and region from user registration information; The behavior feature modeling sub-module (220) constructs user behavior patterns by using time series analysis algorithms, including: The short-term interest model (221) calculates real-time interest weights based on behavior data in the past 7 days; The long-term preference model (222) uses an LSTM neural network to mine behavior patterns exceeding 30 days; The consumption ability evaluation sub-module (230) constructs a consumption power index by combining historical order amounts, payment methods, and coupon usage frequencies; The scenario perception sub-module (240) identifies the scenario characteristics of users through geographical location data and device sensors.

5. An intelligent advertising delivery system based on big data analysis according to claim 4, characterized in that, The behavior feature modeling sub-module (220) adopts a dynamic decay algorithm: Automatically reduces the priority of the label when the user has no relevant behavior for 24 consecutive hours.

6. An intelligent advertising delivery system based on big data analysis according to claim 1, wherein, The advertisement intelligent matching module (300) includes: The feature engineering sub-module (310) maps user features and advertisement features to the same vector space, where: The dimension of the user feature vector ≥ 256; The dimension of the advertisement feature vector ≥ 128; The deep matching sub-module (320) adopts a two-tower neural network architecture, including: The user feature tower (321) is composed of three layers of fully connected networks, and the activation function is LeakyReLU; The advertisement feature tower (322) integrates CNN for processing image advertisements and BERT for processing text advertisements; The multi-objective optimization sub-module (330) balances the Pareto front solution sets of the advertiser's ROI, user experience, and platform revenue.

7. An intelligent advertising delivery system based on big data analysis according to claim 6, characterized in that, The training process of the deep matching sub-module (320) includes: Construct a training set containing 100 million user-advertisement interaction records; Adopt a contrast loss function, and the negative sample sampling ratio is 1:5; Update the model parameters through the dynamic weight averaging algorithm, and the learning rate is set to 0.001; Perform an online model hot update every 6 hours.

8. An intelligent advertising delivery system based on big data analysis according to claim 1, characterized in that, The real-time bidding decision module (400) includes: The dynamic bidding sub-module (410) adjusts the bid according to the user value coefficient and the quality of the ad slot; The frequency control sub-module (420) uses a sliding window algorithm to limit the number of user exposures, and the window period is 24 hours; The cross-platform coordination sub-module (430) synchronously processes the bidding requests of multiple advertising trading platforms (DSPs), and the priority strategies include: High-value users are given priority to display high-price advertisements; New ad materials are given priority to obtain test traffic; Dynamically allocate budgets according to the platform's historical performance.

9. An intelligent advertising delivery system based on big data analysis according to claim 8, characterized in that, The frequency control sub-module (420) realizes: Establish a user-advertisement exposure matrix to record the number of exposures of each user to various types of advertisements; Suspend the display of similar advertisements for 12 hours when the fatigue level exceeds 0.

8.

10. An intelligent advertising delivery system based on big data analysis according to claim 1, characterized in that, The effect evaluation and optimization module (500) includes: The multi-dimensional analysis sub-module (510) calculates 12 core indicators such as the advertisement conversion rate, earnings per thousand impressions (eCPM), and user retention rate; The anomaly detection sub-module (520) uses the isolation forest algorithm to identify abnormal placement events, and the detection response time < 500ms; The policy iteration sub-module (530) dynamically adjusts the parameters of an intelligent advertising placement system based on big data analysis based on a deep reinforcement learning framework, including: State space: contains more than 50 real-time placement features; Action space: adjust the matching threshold, bid coefficient, and frequency limit parameters; The A / B test sub-module (540) runs multiple groups of policies in parallel and compares the effect differences, and the decision-making cycle is 6 hours.