Data promotion system and method of mobile internet
By using a distributed data acquisition framework, a lightweight BERT model, and GAN technology, the problem of noisy data processing in mobile internet data promotion systems was solved, enabling real-time capture of user intent and improvement of data quality, thereby enhancing the accuracy and efficiency of the promotion system.
Patent Information
- Application Number
- CN202510915618.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-03
- Publication Date
- 2025-11-14
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Traditional mobile internet data promotion systems lack adaptive purification mechanisms when dealing with noisy data in distributed collection environments. This leads to distorted behavioral pattern analysis during profile building, a disconnect between semantic information and behavioral data, an inability to accurately understand user intent, and insufficient accuracy in matching promotion templates.
Data processing is performed using a distributed data acquisition framework, a lightweight BERT model, and a generative adversarial network (GAN). Data is purified through real-time computation and adaptive adversarial learning. Combined with semantic perception and behavior analysis, dynamic semantic tags are generated to optimize the push strategy.
It enables real-time capture of user intent and improves data quality, eliminates noisy data, and enhances the accuracy of user profiles and promotion effectiveness, upgrading the promotion system from passive response to proactive prediction.
Smart Images
Figure CN120950774A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data promotion technology, and in particular to a data promotion system and method for mobile internet. Background Technology
[0002] Internet data promotion is a technical system that collects and analyzes user behavior data, builds user profiles, and pushes content. In mobile application scenarios, it collects behavioral data such as clicks and browsing through tracking technology, processes the data to generate user tags, and then sends promotional content through push channels based on the tag matching template. Finally, it optimizes the promotion effect through review.
[0003] However, traditional promotion systems rely on conventional cleaning methods for behavioral data processing, which are insufficient to handle complex quality issues such as abnormal operation records and data deviations. In particular, they lack adaptive cleaning mechanisms for noisy data caused by uneven server load and network fluctuations in distributed collection environments. This results in distorted behavioral pattern analysis during profile construction. At the same time, semantic information and behavioral data are processed separately, and no cross-modal collaborative optimization mechanism has been established. When there are multiple semantic ambiguities with multiple meanings, it is impossible to verify the accuracy of semantic understanding through behavioral data, nor can semantic constraints guide the generation of behavioral data, leading to insufficient accuracy in promotion template matching.
[0004] Accordingly, this application proposes a data promotion system and method for mobile Internet. Summary of the Invention
[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a mobile internet data promotion system and method.
[0006] To achieve the above objectives, the present invention adopts the following technical solution:
[0007] The mobile internet data promotion system includes the following modules:
[0008] Basic data processing module
[0009] Build a distributed data acquisition framework, collect user behavior data, format it, and store it in the cloud with encryption.
[0010] General portrait building module
[0011] Based on user behavior data, a real-time computing framework is used to aggregate multi-dimensional data, supporting sliding time window calculation and incremental updates; through feature engineering and machine learning algorithms, basic profile dimensions such as population attribute tags, behavior tags, and interest tags are generated, supporting dynamic adjustment of tag weights and hierarchical management of profiles.
[0012] Templated content push module
[0013] It stores promotional templates in various formats such as images, videos, and H5, and supports template version management and A / B testing; it uses algorithms such as collaborative filtering and deep learning based on user profiles to perform content matching, and optimizes push strategies by combining real-time scenario data to achieve tens of millions of concurrent push scheduling.
[0014] Fixed channel transmission module
[0015] Select the appropriate push channel based on the user's device type, operating system version, and network status; support multi-channel redundancy backup and intelligent switching; set personalized push parameters for different channels, including push time, frequency, priority, notification bar style, etc.; optimize delivery strategy through real-time monitoring and feedback mechanism of push effect.
[0016] Manual review and optimization module
[0017] Visualization tools are used to display promotional data and calculate metrics such as click-through rate and conversion rate. Based on the data analysis results, promotional templates, push strategies, and system parameters are adjusted to achieve iterative optimization of promotional effectiveness.
[0018] Preferably, the basic data processing module uses Kubernetes for container orchestration and dynamically allocates collection tasks through a load balancing algorithm.
[0019] The mobile internet data promotion method, implemented by the basic data processing module of the mobile internet data promotion system, includes the following steps:
[0020] S1. Initial Deployment: Build a distributed framework based on Kubernetes, configure a high-availability architecture with multiple availability zones, and support cross-regional traffic scheduling; allocate hybrid computing resources such as CPU / GPU / memory, ensure the performance of critical business through resource isolation technology, and dynamically allocate collection tasks using load balancing algorithms;
[0021] S2, Behavior Capture
[0022] By capturing user's regular behavior data and text input data through deep data tracking technology, it supports full-link data tracking and attribution analysis; it adds multi-dimensional identifiers such as device ID, session ID, and event ID to the collected data, and uses a real-time data stream monitoring system to dynamically monitor data quality.
[0023] S3, Formatting
[0024] It uses JSONSchema to convert heterogeneous data formats, supports custom data mapping rules and field-level validation; it develops industry-specific data templates, establishes a quality scoring mechanism that includes indicators such as completeness, accuracy, and consistency, and removes low-quality data through an automated data cleaning pipeline.
[0025] S4, cloud storage
[0026] The AES-256 encryption algorithm is used to perform end-to-end encryption of data, supporting regular key rotation and hierarchical access control; a distributed data lake architecture is built, and the cost of cold and hot data is optimized through a hierarchical storage strategy, a cross-module data sharing mechanism is established, and second-level data subscription and distribution are supported.
[0027] S5, Dynamic Semantic Awareness
[0028] The lightweight BERTTiny model is deployed to process text data. The model size is optimized by quantization compression technology, semantic vectors are extracted and associated with user behavior data to generate dynamic semantic tags. The millisecond-level response to changes in user intent is achieved by calculating cosine similarity.
[0029] S6, Adaptive Adversarial Learning
[0030] A generative adversarial network (GAN) is constructed, and a semantic constraint loss function is designed. The GAN is trained using semantic data as input. Reinforcement learning is combined to dynamically adjust the parameters of the generator and discriminator, optimize the behavioral pattern matching degree of the generated data, and improve the data quality through multiple rounds of iteration.
[0031] Preferably, in the initial deployment step, the data collection tasks are dynamically allocated using a load balancing algorithm.
[0032] Preferably, in the behavior capture step, contextual information of text input is captured using tracking technology.
[0033] Preferably, in the dynamic semantic perception step, changes in user intent are determined by cosine similarity.
[0034] Preferably, in the adaptive adversarial learning step, the behavioral pattern matching degree of the generated data is optimized by combining semantic information.
[0035] The present invention has the following beneficial effects:
[0036] By deploying a lightweight BERT model to extract semantic vectors from text and calculating similarity, real-time capture of changes in user intent is achieved, generating data records with semantic tags. This step solves the problem of delayed intent understanding caused by traditional systems relying solely on behavioral data, providing precise semantic guidance for data collection, improving the matching degree between data and users' real needs, and laying a semantic understanding foundation for subsequent personalized promotion.
[0037] By constructing a GAN network and training it adversarially with semantic data as input, the generator and discriminator engage in a game of adversarial analysis to purify abnormal data, and the generated data is optimized by combining semantic information. This step addresses the problem that traditional data cleaning methods struggle to handle complex behavioral distributions, eliminating low-quality data and correcting behavioral pattern biases caused by noise, thus providing data support that more closely reflects real-world behavior for user profiling.
[0038] By combining dynamic semantic perception with adaptive adversarial learning to form a cross-modal model, semantic vectors serve as semantic logic for the data generated as training conditions for GANs. The purified data then feeds back to improve the robustness of the semantic model. This mechanism overcomes the limitations of single-modal processing, achieving joint optimization in areas such as semantic ambiguity resolution and behavioral data completion. It drives the system to upgrade from a passive response based on statistical laws to an active prediction mode of joint semantic-behavioral reasoning. Attached Figure Description
[0039] Figure 1 This is a flowchart of the subsequent options for distributed data collection in mobile internet proposed in this invention;
[0040] Figure 2 A flowchart showing the options for enabling GAN purification in the mobile internet data promotion method proposed in this invention;
[0041] Figure 3 This is a scatter plot showing the response time of the mobile internet data promotion method proposed in this invention at different time periods. Detailed Implementation
[0042] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.
[0043] Example 1:
[0044] Basic data processing module
[0045] S1. Initial Deployment
[0046] A distributed data acquisition framework was built in the mobile application backend, using Kubernetes for container orchestration to achieve clustered management of data acquisition microservices. The data acquisition frequency was set to 60 times per minute, determined through a comprehensive evaluation of server capacity, network bandwidth, and data timeliness. Using a load balancing algorithm, acquisition tasks were dynamically allocated based on the real-time resource status of each server node, such as CPU utilization and memory usage. For example, when the CPU utilization of a node exceeds 80%, new tasks are assigned to other nodes with lower loads to maintain system stability. Simultaneously, the data acquisition scope was clearly defined in the configuration file, covering all behavioral events within the app, including page navigation, button clicks, and content browsing, laying a solid foundation for subsequent data processing.
[0047] S2, Behavior Capture
[0048] Using event tracking technology, multi-dimensional user behavior data is collected within mobile applications, including actions such as clicking on product detail pages, adding items to the shopping cart, and submitting orders, as well as information such as page dwell time and scroll distance, to construct a data space D = {X1, X2, ..., X...}. m To ensure data quality, data integrity metrics are defined. Regularly review the collected data. Once I D If the data rate drops below 80%, the data replenishment mechanism will be triggered immediately to re-collect the missing data and ensure that the data is complete and usable.
[0049] S3, Formatting
[0050] The JSON Schema specification is adopted to uniformly convert the raw data into standard JSON format. During the conversion process, data validation tools are used to strictly verify the data to ensure that it conforms to the predefined pattern. Taking user click event data as an example, it is explicitly required that it must include fields such as event type, occurrence time, and page URL. If any field is missing, the data will be deemed invalid and corrected to ensure that the data format is standardized and consistent, facilitating subsequent storage and analysis.
[0051] S4, cloud storage
[0052] Data is encrypted using the AES-256 encryption algorithm. The key is generated using a secure key generation function K = Genkey(seed), where seed is a random seed. The encryption formula is E = Encrypt(D,K), ensuring data transmission and storage security. Encrypted data is uploaded to Alibaba Cloud OSS or AWS S3 storage buckets via HTTPS. During upload, a chunked upload method is used, splitting large files into multiple chunks for transmission, supporting resumeable uploads to ensure complete data upload.
[0053] General portrait building module
[0054] Based on the collected user behavior data, the average browsing time per user is calculated through statistical analysis. Where T i The system calculates the duration of a user's i-th page visit and tracks the percentage of visits to different page types (e.g., news, e.g., shopping). Based on preset thresholds, basic user profile tags are generated. For example, users with an average daily browsing time exceeding 30 minutes are tagged as "high-frequency users," and users whose shopping page visits account for over 60% of their visits are tagged as "users who prefer shopping pages." This provides user characteristic data for subsequent promotions.
[0055] Templated content push module
[0056] The promotion template set M = {m1, m2, ..., m} is stored in the Redis cache. k Each template includes fields such as title, body text, image links, and applicable scenarios. Based on the user's basic profile tags, a selection function f is used. select (L)=m j Template matching. If a user's tag includes "user who prefers shopping pages", templates featuring product promotions will be prioritized; if the user is a "frequent browser", templates with more recommended content will be pushed, achieving initial personalized push notifications.
[0057] Fixed channel transmission module
[0058] The function `Push(message, channel)` uses Firebase Cloud Messaging (FCM) or Huawei Push Service (HMS) to automatically select the push channel to send promotional content based on the user's device type (Android / iOS). Parameters such as time and frequency can be set during push notifications, for example, no pushes will be sent during non-working hours (22:00-07:00), and the same type of message will be pushed to the same user no more than twice per day, balancing promotional effectiveness and user experience.
[0059] Manual review and optimization module
[0060] In tools like Tableau or Power BI, you can visualize the click-through rate calculated from the points of a database query. Conversion rate To evaluate the effectiveness of the promotion, reports are manually analyzed. If a template has a low click-through rate, the design is considered for replacement; if the conversion rate is high during a certain period, the number of pushes during that period is increased in subsequent promotions to gradually optimize the promotion effect.
[0061] Example 2:
[0062] S1. Initial Deployment
[0063] A distributed data acquisition framework was built in the mobile application backend, using Kubernetes for container orchestration, and configured to acquire data 60 times per minute. A lightweight data preprocessing service was additionally deployed on the server side to receive raw text data uploaded from mobile devices, perform preliminary cleaning and format standardization, reducing the burden on subsequent processing. Simultaneously, server memory allocation and computing resource scheduling strategies were optimized to adapt to the BERTTiny model.
[0064] S2, Behavior Capture
[0065] Building upon conventional behavioral data collection, the focus is on capturing user-input text-related operation events, such as text data collection commands triggered by entering text in the search box, submitting comments, and sending in-site messages. For each text input action, metadata such as input time and input location (page and interactive components) are recorded, forming a complete data record together with the text content. Furthermore, to meet semantic analysis needs, the collection of text input context information is added, such as the user's browsing pages and operational behaviors before entering text.
[0066] S3, Formatting
[0067] In addition to converting behavioral data to JSON format, specific formatting specifications were established for text data, requiring the inclusion of fields such as original text content, text type (search terms / comments, etc.), and encoding format. Regular expressions and semantic validation tools were used to verify the legality and integrity of the text data, ensuring data quality. Semantic vector data was also formatted uniformly to enable it to be processed collaboratively with other data.
[0068] S4, cloud storage
[0069] Data is encrypted using the AES-256 encryption algorithm. For key management, a Hardware Security Module (HSM) is introduced to store the keys, enhancing key security. Encrypted data is uploaded to a cloud database via HTTPS, and a data version management mechanism is established for easy traceability and analysis. To facilitate the retrieval and access of semantically relevant data, a data index based on semantic keywords is created.
[0070] S5, Dynamic Semantic Awareness
[0071] Deploying the compressed and optimized BERTTiny lightweight pre-trained language model on mobile devices allows for adaptation to limited device resources. It analyzes user-input text such as search keywords, product reviews, and customer service chat logs in real time, processing the input text S to obtain the semantic vector υ = BERT. Tiny (S). By cosine similarity Calculate the similarity of semantic vectors from different texts to determine changes in user intent. For example, if a user enters "phone" followed by "smartphone," a high similarity indicates a consistent intent; conversely, if they enter "phone accessories," a low similarity indicates a change in intent, providing direction for accurate data collection.
[0072] General portrait building module
[0073] Based on the collected user behavior data and semantic analysis results, the average browsing time per user is calculated. Based on the percentage of page views, new semantically related tags are added. For example, if a user frequently inputs semantic vectors related to "photography" and reaches a certain threshold, combined with their browsing behavior on photography-related pages, a "deep photography interest user" tag is generated, enriching the user profile dimensions and providing more accurate user characteristics for subsequent promotions.
[0074] Templated content push module
[0075] The promotion template set M = {m1, m2, ..., m} is stored in the Redis cache. k Each template, in addition to the original fields such as title, body text, image links, and applicable scenarios, adds semantic adaptation tags. Based on user profile tags and tags derived from semantic analysis, the tags are selected using the function f. select (L)=m j Matching templates. For example, for "users with a deep interest in photography," prioritize pushing templates that are semantically appropriate, such as introductions to professional photography equipment and sharing of photography tips, to further enhance the personalized push effect.
[0076] Fixed channel transmission module
[0077] Define a message push function that uses Firebase Cloud Messaging (FCM) or Huawei Push Service (HMS) to automatically select the push channel to send promotional content based on the user's device type (Android / iOS). Given that semantic awareness may capture potential information about user message receiving preferences (such as user expressions of preference for certain types of push notifications in chat), in addition to setting conventional parameters such as time and frequency, the priority of the push channel is dynamically adjusted based on semantic analysis results to improve message delivery effectiveness.
[0078] Manual review and optimization module
[0079] In tools like Tableau or Power BI, data such as click counts, impressions, and conversions from database queries can be visualized, as shown below:
[0080]
[0081] This is used to evaluate the effectiveness of the promotion. Simultaneously, the analysis of semantically relevant metrics is increased, such as click-through rate based on semantic tags and conversion rates for different semantic templates. When manually analyzing reports, if a low click-through rate is found for a particular semantically compatible template, not only is a redesign considered, but further analysis of semantic matching is conducted to adjust semantically related promotional strategies, achieving more refined promotional optimization.
[0082] Example 3
[0083] S1. Initial Deployment
[0084] A distributed data acquisition framework was built in the mobile application backend, using Kubernetes for container orchestration, and configured with a data acquisition frequency of 60 times per minute. Dedicated GPU computing resources were reserved on the cloud server for subsequent training and inference of the Generative Adversarial Network (GAN), ensuring that computing performance met the requirements. Simultaneously, a data monitoring system was built to monitor various indicators during data acquisition, transmission, and storage in real time, providing a stable data infrastructure environment for GAN training.
[0085] S2, Behavior Capture
[0086] User behavior data is collected using event tracking technology. During the data collection process, the data undergoes initial screening to remove obviously erroneous or duplicate records, reducing the transmission and storage of invalid data. Simultaneously, a data source identifier and collection timestamp are added to each data entry to facilitate subsequent data traceability and analysis. Furthermore, system environment information during data collection, such as server load and network status, is recorded to analyze the correlation between data quality and environmental factors during GAN training.
[0087] S3, Formatting
[0088] The raw data is converted to standard JSON format using the JSON Schema specification, with detailed data format templates developed for different types of behavioral data (such as clicks, views, purchases, etc.). Data validation tools are used to rigorously validate the converted data to ensure it conforms to the predefined patterns; any abnormal data is recorded and flagged. To accommodate the specific data format requirements of GAN training, the data is normalized and standardized to conform to the model input specifications.
[0089] S4, cloud storage
[0090] Data is encrypted using the AES-256 encryption algorithm, and a distributed storage architecture is employed to store data across multiple storage nodes, improving the reliability and availability of data storage. A data indexing mechanism is established, creating indexes based on key data attributes (such as user ID, behavior type, and time) to accelerate data query and retrieval speeds. Simultaneously, a separate storage area is allocated for GAN training data, and the training data is regularly backed up and cleaned to ensure the validity of the training data and the rational utilization of storage resources.
[0091] S6, Adaptive Adversarial Learning
[0092] A Generative Adversarial Network (GAN) is built on a cloud server, consisting of a generator G and a discriminator D. The generator simulates user behavior to generate fake data, and the discriminator distinguishes between real and fake data by optimizing the objective function. Adversarial training. During training, the generator adjusts its parameters to generate more realistic data, while the discriminator improves its discriminative ability, thereby identifying and eliminating low-quality or biased data. For example, if the generator generates abnormal behavior data, the discriminator's feedback prompts it to adjust and purify the data. During training, the GAN's training parameters, such as the learning rate and number of iterations, are dynamically adjusted based on changes in data distribution to improve the model's adaptability to complex user behaviors.
[0093] General portrait building module
[0094] Based on the collected user behavior data and semantic analysis results, the average browsing time per user is calculated. When calculating the percentage of page views, the data undergoes secondary filtering and analysis. Because the data purified by GANs better reflects users' true behavioral patterns, the generation of basic user profile tags is based on more accurate data characteristics, making the definitions of tags such as "high-frequency browsing users" and "users who prefer shopping pages" more closely match actual user behavior and improving the accuracy of user profiles.
[0095] Templated content push module
[0096] The promotion template set M = {m1, m2, ..., m} is stored in the Redis cache. k Each template includes fields such as title, body text, image links, and applicable scenarios. Based on the user's basic profile tags, a selection function f is used. select (L)=m j Template matching. Considering that the data processed by GAN may reveal potential user behavioral characteristics, the template matching process incorporates analysis of these potential characteristics, prioritizing the push of templates related to users' potential interests to uncover more user needs and improve promotion effectiveness.
[0097] Fixed channel transmission module
[0098] The function `Push(message, channel)` uses Firebase Cloud Messaging (FCM) or Huawei Push Service (HMS) to automatically select the appropriate push channel to send promotional content based on the user's device type (Android / iOS). Parameters such as time and frequency can be set during push notifications; for example, pushes will not be sent during non-working hours (22:00-07:00), and the same type of message will not be pushed to the same user more than twice per day. Based on GAN analysis of user behavior, the system predicts the likelihood of users receiving messages at different times and dynamically adjusts the push time to improve message reception rates.
[0099] Manual review and optimization module
[0100] In tools like Tableau or Power BI, data such as click counts, impressions, and conversions from database queries can be visualized, as shown below:
[0101]
[0102] This is used to evaluate the promotion effect. Simultaneously, metrics during the GAN training process are analyzed, such as changes in the loss functions of the generator and discriminator, and the similarity between generated and real data. When manually analyzing reports, these metrics are used to assess the impact of data cleansing on the promotion strategy. If poor GAN training is found to affect promotion, the GAN training strategy or parameters are adjusted to achieve synergistic optimization between the promotion strategy and data processing.
[0103] Example 4
[0104] Basic data processing module
[0105] S1. Initial Deployment
[0106] A distributed data acquisition framework was built in the mobile application backend, using Kubernetes for container orchestration, and configured with a data acquisition frequency of 60 times per minute. Lightweight data preprocessing services and GPU computing resources were deployed simultaneously on the server side to meet the computational needs of dynamic semantic awareness and adaptive adversarial learning, respectively. Server resources were rationally planned to avoid conflicts in resource usage between the two innovative functions. A unified data management platform was also built to achieve centralized management of different types of data and processing workflows.
[0107] S2, Behavior Capture
[0108] Building upon the collection of routine behavioral data, the focus is on capturing user input text-related operation events. Simultaneously, the collected data undergoes initial screening and preprocessing. Multi-dimensional identification information is added to each data entry, including the data source, data type, and associated user actions, facilitating subsequent data integration and analysis. During data collection, data quality is monitored in real-time. Combining semantic awareness and adversarial learning requirements for data quality, the collection strategy is dynamically adjusted to ensure that the collected data better supports subsequent innovative features.
[0109] S3, Formatting
[0110] The raw data is converted to standard JSON format using the JSON Schema specification. Detailed format templates are developed for both text and behavioral data, and rigorous data validation is performed. A data quality scoring mechanism is established to score each data point based on indicators such as completeness, accuracy, and consistency, facilitating subsequent data filtering and use. Corresponding format standards are also developed for semantic vector data and GAN-processed data, enabling unified management and efficient processing of different data types.
[0111] S4, cloud storage
[0112] Data is encrypted using the AES-256 encryption algorithm, and a distributed storage and data redundancy backup strategy is employed to ensure data security and reliability. A data lake storage architecture is constructed to uniformly store and manage data of different types and sources, providing a foundation for subsequent data analysis and mining. A data sharing mechanism is established to enable dynamic semantic awareness and adaptive adversarial learning modules to easily obtain and use the required data, while ensuring data security and consistency.
[0113] S5, Dynamic Semantic Awareness
[0114] A compressed and optimized lightweight pre-trained BERTTiny language model is deployed on mobile devices to analyze user input text in real time, obtain semantic vectors, calculate similarity, and capture changes in user intent. The semantic analysis results are correlated with user behavior data to generate behavior data records with semantic tags. Simultaneously, the semantic analysis data is transmitted to an adaptive adversarial learning module, providing it with richer semantic information to aid in data cleansing and model training.
[0115] S6, Adaptive Adversarial Learning
[0116] A Generative Adversarial Network (GAN) is built on a cloud server. Adversarial training optimizes the objective function, cleanses the data, and enhances the system's ability to understand complex user behaviors. Dynamically semantically processed data is used as training data input for the GAN, further improving data quality and model performance. During training, semantic information is incorporated to optimize the generated data, ensuring that it not only conforms to user behavior patterns but also better matches user semantic intent.
[0117] General portrait building module
[0118] Based on collected user behavior data, semantic analysis results, and data purified by GAN, the average browsing time per user is comprehensively calculated. Detailed statistics on the percentage of visits to various page types are compiled and analyzed to generate multi-dimensional user profile tags. Semantic tags are used to determine user interests and preferences. For example, if a user frequently inputs semantic vectors related to "outdoor sports" and spends a significant amount of time browsing outdoor sports equipment pages, and the GAN-purified data also supports this behavioral pattern, then the tag "deep outdoor sports enthusiast" is generated. User behavior pattern tags are then corrected based on the GAN-purified data, eliminating inaccurate tags caused by data noise or abnormal behavior. This makes the profile more accurately reflect the user's true behavior and interests, providing comprehensive and accurate user characteristic data for promotion.
[0119] Templated content push module
[0120] The promotion template set M = {m1, m2, ..., m} is stored in the Redis cache. k Each template, in addition to its original fields, adds semantic-behavioral dual-adaptation tags. Based on the user's composite profile tags, a selection function f is used. select (L)=m j Matching templates. For example, for "deep outdoor sports enthusiasts," prioritize pushing templates that include recommendations for professional outdoor sports equipment and match their sports interests and habits, such as "Limited-time special offer on professional hiking boots to help you conquer every peak," and pair them with outdoor sports guides, sports community invitations, etc., to increase users' interest and participation in the promoted content and achieve more accurate and personalized content delivery.
[0121] Fixed channel transmission module
[0122] The function `Push(message, channel)` automatically selects the appropriate push channel to send promotional content based on the user's device type (Android / iOS), leveraging Firebase Cloud Messaging (FCM) or Huawei Push Service (HMS). Combining user message receiving preferences obtained through dynamic semantic awareness with user behavior predictions from adaptive adversarial learning, it not only sets conventional parameters such as time and frequency but also further optimizes the selection and priority of push channels. For example, if a user has expressed a preference for SMS push notifications in a chat, and the adversarial learning model predicts that the user is more likely to check SMS messages during a certain time period, then important promotional information will be prioritized for push via the SMS channel during that time period, improving message reach and user reception rate.
[0123] Manual review and optimization module
[0124] In tools like Tableau or Power BI, data such as click counts, impressions, and conversions retrieved from the database can be visualized, as shown below:
[0125]
[0126] This is used to evaluate the promotion's effectiveness. Simultaneously, in-depth analysis of semantic perception and adversarial learning-related metrics is conducted, such as click-through rate based on semantic tags, conversion rate of different semantic-behavior matching templates, changes in the loss function during GAN training, and the similarity between generated and real data. When manually analyzing reports, these multi-dimensional metrics are used to comprehensively evaluate the impact of the dual-innovation features on the promotion strategy. If a low click-through rate is found for a particular semantic-behavior matching template, not only is redesigning considered, but also in-depth analysis of semantic matching degree, behavioral fit, and data quality is performed. This involves adjusting semantic analysis model parameters, adversarial learning training strategies, or promotion content design to achieve deep synergistic optimization of the promotion strategy with data processing and model training.
[0127] It should be noted that in the comparative embodiments, data promotion is based solely on basic user behavior data and uses fixed templates and channels. The data collection and processing methods are simple, resulting in poor data quality and low processing efficiency. Furthermore, relying on preset simple rules for data filtering and promotion content matching lacks advanced technologies such as deep learning. The limitations of these rules lead to lower accuracy and efficiency in data processing, as shown in Table 1.
[0128] Table 1: Quantitative Comparison and Evaluation Table of the Implementation Effect of Mobile Internet Data Promotion Plans
[0129]
[0130] Specifically, Comparative Example 1 uses traditional push notifications, whose data collection relies on simple basic behavior records. The lack of verification mechanisms in transmission and storage leads to data loss and frequent errors, resulting in an integrity rate of only 78.6%. Comparative Example 2 is based on a rule engine. Although it has filtering rules, these rules are rigid and lack flexibility. When faced with complex user behaviors and abnormal operations, it struggles to comprehensively collect data, improving integrity to only 82.1%. In contrast, Example 1's distributed collection framework and supplementary collection mechanism ensure the comprehensiveness and accuracy of data collection from an architectural perspective, achieving an improvement of nearly 10% compared to the comparative example. Example 2 introduces dynamic semantic awareness, deeply mining the intent behind user text operations and supplementing the collection of related data, further enhancing integrity. Example 3's GAN technology accurately identifies and corrects abnormal data through adversarial training, eliminating data noise. Finally, Example 4 integrates two innovative technologies, providing refined control over the entire data collection, transmission, and storage chain, optimizing processing at each stage to achieve 96.5% integrity. Compared to Comparative Example 1, data integrity is improved by nearly 23%, laying a solid foundation for accurate analysis and efficient promotion.
[0131] Specifically, Comparative Examples 1 and 2, due to their rudimentary technologies, struggle to identify complex abnormal data patterns. Comparative Example 1 lacks a dedicated filtering mechanism, and the preset rules of Comparative Example 2's rule engine cannot cover all abnormal scenarios, resulting in filtering rates of only 76.3% and 80.4% respectively. A large amount of noisy data remains, severely impacting the accuracy of subsequent analysis and generalization. The format validation and basic screening in Example 1 can only handle format-level anomalies, with limited filtering efficiency. Example 3 introduces GAN technology, which, through the game between the generator and discriminator, learns the distribution characteristics of real data, accurately locates and removes abnormal data deviating from normal patterns, increasing the filtering rate to 91.5%. Example 4, after incorporating semantic perception, utilizes user intent information to assist in anomaly detection. For example, when a user's semantic expression is "searching for a mobile phone," but behavioral data shows frequent clicks on clothing pages, the contradiction between semantics and behavior allows for more accurate anomaly identification, achieving a filtering rate of 94.3%, nearly 24% higher than Comparative Example 1, significantly improving data purity.
[0132] Specifically, Comparative Examples 1 and 2 have rudimentary processing architectures. Comparative Example 1 lacks efficient algorithms and parallel processing capabilities, while Comparative Example 2's serial processing method for rule judgments experiences a sharp drop in efficiency as the data volume increases. This results in processing speeds of only 110.7 MB / h and 135.2 MB / h respectively, insufficient to meet the demands of large-scale data processing. Example 1's distributed architecture initially improves processing efficiency, while Example 2 optimizes data acquisition and processing flows through semantic awareness, reducing invalid data interference and increasing processing speed. Example 3 utilizes GPUs to accelerate GAN training and inference, combined with data index optimization, significantly improving data processing efficiency. Example 4, through resource integration, the introduction of automated tools, and high-speed network optimization, constructs a highly efficient data processing pipeline. For example, automated format conversion tools reduce manual processing time by approximately 40%, and high-speed networks improve data transmission efficiency by over 30%, ultimately achieving a processing speed of 265.8 MB / h, 2.4 times that of Comparative Example 1. This enables rapid response to massive data processing needs and supports large-scale business deployment.
[0133] Specifically, Comparison Example 1 and Example 2 exhibit rigid push strategies. Comparison Example 1 uses fixed templates and channels, failing to consider users' personalized needs and real-time status; Comparison Example 2 suffers from slow rule matching processes, resulting in response times of 4.5s and 3.8s respectively, making users highly susceptible to churn and significantly reducing promotional effectiveness. Example 1 optimizes the infrastructure, shortening some process times. Example 2 uses semantic precision to match user needs, reducing the time spent pushing invalid content. Example 3 plans push timing and content more rationally based on purified user behavior data. Example 4 integrates semantic intent and behavioral prediction to dynamically adjust the push strategy. For example, by analyzing users' active periods of semantic expression and usage habits in behavioral data, pushes are sent when users are most likely to receive information, while optimizing channel priorities, reducing response time to 2.3s, nearly 50% shorter than Comparison Example 1, significantly improving user experience and promotional immediacy, ensuring promotional content reaches users promptly and increasing conversion rates.
[0134] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A mobile internet data promotion system, characterized in that, Includes the following modules: Basic data processing module Build a distributed data acquisition framework, collect user behavior data, format it, and store it in the cloud with encryption. General portrait building module Based on user behavior data, a real-time computing framework is used to aggregate multi-dimensional data, supporting sliding time window calculation and incremental updates; through feature engineering and machine learning algorithms, basic profile dimensions such as population attribute tags, behavior tags, and interest tags are generated, supporting dynamic adjustment of tag weights and hierarchical management of profiles. Templated content push module It stores promotional templates in various formats such as images, videos, and H5, and supports template version management and A / B testing; it uses algorithms such as collaborative filtering and deep learning based on user profiles to perform content matching, and optimizes push strategies by combining real-time scenario data to achieve tens of millions of concurrent push scheduling. Fixed channel transmission module Select the appropriate push channel based on the user's device type, operating system version, and network status; support multi-channel redundancy backup and intelligent switching; set personalized push parameters for different channels, including push time, frequency, priority, notification bar style, etc.; optimize delivery strategy through real-time monitoring and feedback mechanism of push effect. Manual review and optimization module Visualization tools are used to display promotional data and calculate metrics such as click-through rate and conversion rate. Based on the data analysis results, promotional templates, push strategies, and system parameters are adjusted to achieve iterative optimization of promotional effectiveness.
2. The mobile internet data promotion system according to claim 1, characterized in that, The basic data processing module uses Kubernetes for container orchestration and dynamically allocates data collection tasks through a load balancing algorithm.
3. A mobile internet data promotion method, implemented by the basic data processing module of the mobile internet data promotion system as described in claims 1-2, characterized in that, Includes the following steps: S1. Initial Deployment: Build a distributed framework based on Kubernetes, configure a high-availability architecture with multiple availability zones, and support cross-regional traffic scheduling; allocate hybrid computing resources such as CPU / GPU / memory, ensure the performance of critical business through resource isolation technology, and dynamically allocate collection tasks using load balancing algorithms; S2, Behavior Capture By capturing user's regular behavior data and text input data through deep data tracking technology, it supports full-link data tracking and attribution analysis; it adds multi-dimensional identifiers such as device ID, session ID, and event ID to the collected data, and uses a real-time data stream monitoring system to dynamically monitor data quality. S3, Formatting It uses JSONSchema to convert heterogeneous data formats, supports custom data mapping rules and field-level validation; it develops industry-specific data templates, establishes a quality scoring mechanism that includes indicators such as completeness, accuracy, and consistency, and removes low-quality data through an automated data cleaning pipeline. S4, cloud storage The AES-256 encryption algorithm is used to perform end-to-end encryption of data, supporting regular key rotation and hierarchical access control; a distributed data lake architecture is built, and the cost of cold and hot data is optimized through a hierarchical storage strategy, a cross-module data sharing mechanism is established, and second-level data subscription and distribution are supported. S5, Dynamic Semantic Awareness The lightweight BERTTiny model is deployed to process text data. The model size is optimized by quantization compression technology, semantic vectors are extracted and associated with user behavior data to generate dynamic semantic tags. The millisecond-level response to changes in user intent is achieved by calculating cosine similarity. S6, Adaptive Adversarial Learning A generative adversarial network was constructed, a semantically constrained loss function was designed, and adversarial training was performed using semantic data as input. By combining reinforcement learning to dynamically adjust the parameters of the generator and discriminator, the matching degree of behavioral patterns in the generated data is optimized, and the data is purified through multiple rounds of iteration to improve data quality.
4. The mobile internet data promotion method according to claim 3, characterized in that, In the initial deployment step, the data collection tasks are dynamically allocated through a load balancing algorithm.
5. The mobile internet data promotion method according to claim 3, characterized in that, In the behavior capture step, contextual information of text input is captured using tracking technology.
6. The mobile internet data promotion method according to claim 3, characterized in that, In the dynamic semantic perception step, changes in user intent are determined by cosine similarity.
7. The mobile internet data promotion method according to claim 3, characterized in that, In the adaptive adversarial learning step, the behavioral pattern matching degree of the generated data is optimized by combining semantic information.