An ai-based potential traffic consumption user determination method and system

Abstract

The application discloses a kind of potential traffic consumption user determination method and system based on AI, it is related to artificial intelligence big data mining technical field.The method is by collecting user basic attribute, traffic consumption, online behavior, terminal scene, environment correlation multidimensional full data, constructs exclusive feature engineering system after being preprocessed, excavates explicit, implicit, time sequence, scene multidimensional core features;Through improved XGBoost AI model, sample training and parameter optimization are completed, and user traffic consumption potential value is accurately predicted;Potential traffic consumption user is screened in combination with dynamic self-adaptive determination threshold, and user is potential layered and is marked, and simultaneously based on real-time operation data, the dynamic iteration update of model and threshold is realized.The application breaks through the technical limitation of traditional rule screening, solves the problems of single data dimension, low recognition accuracy, rule solidification and inability to adaptively iterate in the prior art, can accurately mine various potential traffic consumption users, greatly improves traffic operation accuracy and conversion efficiency, reduces marketing resource loss, and is suitable for the large-scale traffic refinement operation scene of operator and internet platform.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence user profiling analysis, communication big data mining and traffic operation technology, specifically to an AI-based method and system for identifying potential traffic consumers, applicable to scenarios of accurate identification, demand prediction and refined operation of potential traffic consumers by operators and Internet platforms. Background Technology

[0002] With the rapid development of mobile communication networks and internet services, data traffic has become a core revenue segment for telecom operators and internet platforms. Against the backdrop of saturated existing user base and increasingly fierce market competition, accurately identifying potential data traffic consumers, improving data conversion efficiency, and reducing blind marketing costs have become core operational demands in the industry.

[0003] Existing user traffic mining technologies mostly employ traditional rule matching and simple statistical analysis, relying on shallow structured data such as users' historical traffic consumption frequency, spending amount, and package type for screening and judgment. This approach has several drawbacks: First, it has a single data dimension, relying solely on explicit consumption data while ignoring implicit core data such as user online behavior, scenario preferences, terminal attributes, time-of-day characteristics, and social connections, failing to comprehensively depict users' traffic consumption potential. Second, its judgment logic is rigid, based on manually set fixed thresholds for user screening, resulting in poor adaptability and difficulty in adapting to the dynamic consumption characteristics of different seasons, marketing activities, and user groups. Third, its level of intelligence is low, failing to uncover non-linear correlation features in user behavior data, leading to extremely low accuracy in identifying low-activity, low-spending users with potential for growth, easily missing high-value potential users, and easily filtering out invalid users, resulting in wasted marketing resources. Fourth, it lacks dynamic iteration capabilities; user consumption habits and online scenarios are constantly changing, and traditional static models cannot update judgment rules in real time, leading to a continuous decline in long-term recognition accuracy.

[0004] In existing technologies, some solutions attempt to introduce machine learning models for user identification. However, these solutions generally suffer from problems such as coarse feature selection, high sample noise, fixed weight allocation, and failure to incorporate features specific to traffic consumption scenarios. Furthermore, their generalization ability is weak, resulting in poor identification of niche potential users, dormant / reawakened users, and incremental consumers. Consequently, they cannot meet the needs of large-scale, precise, and dynamic traffic user mining and operation. Therefore, there is an urgent need for an intelligent, high-precision, and dynamically iterative technical solution for identifying potential traffic consumers. Summary of the Invention

[0005] This invention aims to overcome the shortcomings of existing technologies, such as single user identification dimensions, rigid rules, low accuracy, inability to dynamically iterate, and serious resource waste. It provides an AI-based method and system for identifying potential traffic consumers. Through multi-dimensional full-volume data collection, AI deep feature mining, hierarchical modeling and prediction, dynamic threshold iteration, and user potential classification, it achieves accurate, efficient, and dynamic identification of potential traffic consumers, significantly improving the accuracy and conversion efficiency of traffic operations and reducing marketing and maintenance costs. Technical solution

[0006] To achieve the above objectives, the first aspect of the present invention provides an AI-based method for determining potential traffic consumers, employing the following technical solution: A method for identifying potential traffic consumers based on AI includes the following steps: S1. Construct a multi-dimensional user data system, collect all-dimensional raw data of target users, including five core types of data: user basic attribute data, historical traffic consumption data, internet behavior data, terminal scenario data, and environmental correlation data. Clean, deduplicate, and normalize all raw data to obtain a standardized user dataset. S2. Based on a standardized user dataset, construct a traffic consumption-specific feature engineering system. Extract explicit consumption features, implicit behavioral features, temporal change features, and scenario-related features through AI feature mining algorithms. Select highly correlated core features and complete the feature weight initialization configuration. S3. Construct a lightweight traffic potential prediction AI model. Train, validate and optimize the AI ​​model using labeled historical user sample datasets, remove noisy samples, calibrate model parameters, and obtain a trained user traffic potential prediction model. S4. Input the standardized feature data of the target users to be identified into the trained AI model to predict the potential value of user traffic consumption. Combine the dynamic adaptive judgment threshold to screen and determine potential traffic consumption users. S5. Stratify the potential traffic consumers into different levels based on their potential, generate user tag profiles, and dynamically iterate and optimize feature weights and model thresholds based on real-time operational data to achieve dynamic updates in user identification.

[0007] Further, in step S1, the user basic attribute data includes user age, gender, region, occupation, user star rating, package tier, and network access duration; the historical data consumption data includes historical monthly / daily data consumption, data package subscription records, add-on package consumption records, data usage balance, and consumption time period distribution; the internet access behavior data includes internet access duration, type of apps accessed, web browsing preferences, frequency of short video / live streaming usage, online gaming duration, and offline caching behavior; the terminal scenario data includes terminal device model, network standard, mobile / fixed scenario, and commuting / home / outdoor usage scenario; and the environmental association data includes seasonal nodes, holidays, regional network traffic peaks, and marketing activity cycles.

[0008] The preprocessing process specifically includes: filling missing values ​​with mean and nearest neighbor interpolation, removing outliers using the 3σ criterion, removing duplicate data based on the user's unique identifier, and mapping feature data of different dimensions to the [0,1] interval through min-max normalization to obtain a structured dataset with a unified standard.

[0009] Furthermore, in step S2, the explicit consumption characteristics include historical average data consumption, consumption growth rate, package utilization rate, and frequency of data overlay; the implicit behavioral characteristics include frequency of use in high-data-consuming scenarios, proportion of time spent using high-frequency data-consuming apps, and density of fragmented internet access; the temporal variation characteristics include the data consumption fluctuation trends over the past 7 days, 30 days, and 90 days, and the behavioral activity change curve; and the scenario-related characteristics include the increase in data consumption during holidays, the proportion of data consumption in outdoor scenarios, and the increase in data consumption for new apps.

[0010] The Pearson correlation coefficient combined with the random forest feature importance algorithm is used to calculate the correlation between each feature and the incremental traffic consumption. Core features with a correlation threshold greater than a preset value are selected, redundant features and invalid noise features are removed, and the initial weight allocation is completed based on the feature contribution.

[0011] Furthermore, in step S3, the AI ​​model adopts an improved XGBoost multi-class prediction model. The model input is the core feature vector, and the output is the user traffic consumption potential value in the range of 0-1. The sample dataset includes positive samples (historical traffic incremental users, new traffic purchase users, and package upgrade users) and negative samples (users with continuously declining traffic, long-term low consumption, and zero incremental users).

[0012] The model training process includes: dividing the sample dataset into training, validation, and test sets in a 7:2:1 ratio; initial model training using the training set; iterative optimization of model learning rate, tree depth, and iteration count hyperparameters using the validation set; validation of model accuracy, precision, and recall using the test set; and removal of noisy samples with errors exceeding a preset threshold until the overall model accuracy reaches the preset standard.

[0013] Furthermore, in step S4, the dynamic adaptive judgment threshold is calculated and generated in real time based on historical recognition accuracy, current marketing conversion rate, and user traffic data, replacing the traditional fixed threshold; the user potential value is compared with the dynamic threshold, and users whose potential value is greater than or equal to the dynamic threshold are judged as potential traffic consumers.

[0014] Furthermore, in step S5, the potential level is divided into three levels: high potential, medium potential, and low potential, each corresponding to different potential value ranges. Differentiated traffic operation strategies are matched for users at different levels, and daily user consumption conversion data is collected in real time. Based on the gradient descent algorithm, the feature weights and judgment thresholds are dynamically updated to achieve continuous iterative optimization of the model.

[0015] The second aspect of this invention provides an AI-based system for determining potential traffic consumers, used to implement the determination method described in any one of the above, including a data acquisition and preprocessing module, an AI feature engineering module, a model training and optimization module, a user potential prediction module, a user hierarchical labeling module, and a dynamic iterative update module; The data acquisition and preprocessing module is used to collect multi-dimensional user raw data, and complete data cleaning, deduplication, normalization preprocessing, and output a standardized user dataset. The AI ​​feature engineering module is used to extract multi-dimensional traffic consumption features based on a standardized dataset, screen core effective features, and complete feature weight initialization. The model training and optimization module is used to build and train an improved XGBoost potential prediction AI model, complete model hyperparameter optimization and accuracy verification, and output a mature prediction model. The user potential prediction module is used to call the trained AI model to predict the traffic consumption potential value of the user to be identified, and to filter potential traffic consumption users by combining dynamic thresholds. The user stratification and labeling module is used to stratify the potential users after screening them by potential level and generate exclusive user profile tags. The dynamic iterative update module is used to dynamically iteratively optimize feature weights, model parameters, and judgment thresholds based on real-time operational conversion data, so as to achieve adaptive system updates. Beneficial effects

[0016] Compared with the prior art, the present invention has the following significant advantages: 1. Comprehensive data dimensions and high depth of feature mining: This invention constructs a five-category comprehensive user data system, breaking through the limitations of traditional single consumer data. Through AI algorithms, it deeply mines explicit, implicit, time-series, and scenario-based core features, accurately capturing users' potential traffic consumption needs, and solving the problems of single features and one-sided recognition in existing technologies.

[0017] 2. High level of model intelligence and excellent recognition accuracy: The improved XGBoost model is combined with a feature importance screening mechanism, abandoning manual fixed rules, accurately mining non-linear correlation features of data, effectively identifying various potential users such as low-activity high-potential and dormant but waking-up users. The recognition accuracy and conversion rate are far higher than those of traditional rule models.

[0018] 3. Possesses dynamic adaptive iteration capability: It innovatively introduces a dynamic adaptive judgment threshold, and dynamically optimizes model parameters and feature weights by combining real-time operation data and overall traffic data. It can adapt to changes in user consumption during different seasons, holidays, and marketing activities, and solves the defects of traditional models that are static and have poor adaptability.

[0019] 4. Highly applicable and valuable for operation: By segmenting and classifying users based on their potential, differentiated and precise operations can be achieved, significantly reducing the resource waste from blind marketing and greatly improving the conversion efficiency of data package subscriptions, add-on packages, and package upgrades, making it suitable for large-scale data operation scenarios of operators. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Figure 1 This is a schematic diagram illustrating the overall process of the AI-based method for determining potential traffic consumers according to the present invention. Figure 2 The accompanying diagram is a schematic diagram of the overall modular architecture of the AI-based potential traffic consumer user determination system of the present invention. Obviously, the figures described below are only some embodiments of the present invention. Detailed Implementation

[0021] The present invention will be further described in detail below with reference to specific embodiments. Example 1: A method for identifying potential traffic consumers based on AI

[0022] like Figure 1 As shown in the figure, this embodiment provides an AI-based method for identifying potential traffic consumers, applied to the scenario of mining the traffic potential of existing users in telecommunications operators. The specific steps are as follows: S1. Data Collection and Preprocessing: Collect five types of raw data from existing users of the operator: basic attributes, 90-day historical traffic consumption, internet access behavior, terminal scenarios, and environmental correlation. Preprocess the raw data to address missing user network access duration, abnormal traffic consumption peaks, and duplicate log data. Fill in missing values ​​through nearest neighbor interpolation, remove abnormal traffic data using the 3σ criterion, and deduplicate user mobile phone numbers to finally obtain a standardized structured user dataset.

[0023] S2. Feature Engineering Construction: Based on a standardized dataset, 28 initial features are extracted, including 8 explicit consumption features, 12 implicit behavioral features, 4 temporal features, and 4 scene association features. Sixteen core features are selected using Pearson correlation coefficient and random forest algorithms, and 12 redundant features are removed. Initial weights are assigned based on the contribution of each feature to the increase in traffic consumption, with high-frequency short video usage time, outdoor scene proportion, traffic consumption growth rate in the past 30 days, and holiday traffic fluctuations being the core high-weight features.

[0024] S3. AI Model Training and Optimization: 100,000 labeled historical user samples were selected, including 60,000 positive samples and 40,000 negative samples, and divided into training, validation, and test sets in a 7:2:1 ratio. An improved XGBoost model was constructed, with a base learning rate of 0.01, a maximum tree depth of 6, and 1000 iterations. Hyperparameters were optimized iteratively using the validation set to remove sample noise. The final model achieved an accuracy of 96.2% and a recall of 94.8% on the test set, meeting the recognition accuracy requirements.

[0025] S4. User Potential Prediction and Screening: The core feature vectors of the 500,000 existing users to be identified are input into the trained AI model, which outputs a traffic potential value for each user in the range of 0-1. Combining the current monthly traffic growth rate and historical marketing conversion rate, a dynamic judgment threshold of 0.62 is calculated in real time, and users with a potential value ≥ 0.62 are screened as potential traffic consumers, resulting in a total of 82,000 high-potential users.

[0026] S5. User Segmentation and Dynamic Iteration: Users are categorized into three tiers: high potential (potential value 0.8-1.0), medium potential (0.62-0.8), and low potential (0.62-0.7). Dedicated user profile tags are generated for each tier. Premium data packages are pushed to high-potential users, data add-on packages to medium-potential users, and experiential data benefits to low-potential users. User conversion data is collected daily, and feature weights and decision thresholds are dynamically updated to continuously optimize the model. Example 2: A Potential Traffic Consumer Determination System Based on AI

[0027] like Figure 2 As shown, this embodiment provides an AI-based system for identifying potential traffic consumers, used to implement the identification method described in Embodiment 1. The system is deployed on a carrier's big data platform, and the collaborative workflow of each module is as follows: The data acquisition and preprocessing module connects in real time to the operator's user database, internet behavior log database, and terminal information database to complete real-time acquisition and standardized preprocessing of data across all dimensions; the AI ​​feature engineering module automatically completes feature extraction, filtering, and weight updates; the model training and optimization module retrains the model on a weekly basis based on newly added sample data and optimizes hyperparameters; the user potential prediction module completes daily batch prediction and filtering of the potential values ​​of all existing users; the user stratification and labeling module automatically generates user potential tags and synchronizes them to the operation platform; and the dynamic iteration update module monitors conversion data in real time and adaptively adjusts the judgment threshold and feature weights.

[0028] Through practical implementation and verification, this system has improved the accuracy of identifying potential traffic users by more than 35%, the traffic marketing conversion rate by more than 28%, and reduced marketing resource consumption by 30% compared with the traditional rule-based screening model, demonstrating extremely high practical application value.

Claims

1. A method for identifying potential traffic consumers based on AI, characterized in that, Includes the following steps: S1. Construct a multi-dimensional user data system, collect five types of raw data from target users: basic user attribute data, historical traffic consumption data, internet behavior data, terminal scenario data, and environmental correlation data. Clean, deduplicate, and normalize all raw data to obtain a standardized user dataset. S2. Based on the standardized user dataset, construct a traffic consumption-specific feature engineering system, extract explicit consumption features, implicit behavioral features, temporal change features, and scene-related features through AI feature mining algorithms, screen highly correlated core features, and complete the feature weight initialization configuration. S3. Construct a lightweight traffic potential prediction AI model. Use labeled historical user sample dataset to train, validate and optimize the parameters of the AI ​​model, remove noisy samples, and obtain the trained user traffic potential prediction model. S4. Input the standardized core feature data of the target user to be identified into the trained AI model to predict the user's traffic consumption potential value. Combine the real-time generated dynamic adaptive judgment threshold to screen and determine potential traffic consumption users. S5. Stratify the potential traffic consumers into different levels and generate user tag profiles. Dynamically iterate and optimize feature weights and model judgment thresholds based on real-time operational data to achieve dynamic identification and updating of potential traffic consumers.

2. The method for determining potential traffic consumers based on AI according to claim 1, characterized in that, In step S1, the user's basic attribute data includes user age, gender, region, occupation, user star rating, package tier, and network access duration; the historical data consumption data includes historical monthly / daily data consumption, data package subscription records, add-on package consumption records, data usage balance, and consumption time period distribution; the internet behavior data includes internet access duration, types of apps accessed, web browsing preferences, frequency of short video / live streaming usage, online gaming time, and offline caching behavior. The terminal scenario data includes terminal device model, network standard, mobile / fixed scenario, commuting / home / outdoor usage scenario; the environmental association data includes seasonal nodes, holidays, regional network traffic peaks, and marketing activity cycles. The preprocessing process specifically involves: filling missing data values ​​using mean fill and nearest neighbor interpolation, removing abnormal data using the 3σ criterion, deduplicating duplicate data based on the user's unique identifier, and mapping the feature data to the [0,1] interval using min-max normalization to obtain a standardized user dataset with unified dimensions.

3. The method for determining potential traffic consumers based on AI according to claim 1, characterized in that, In step S2, the explicit consumption characteristics include historical average data consumption, consumption growth rate, package utilization rate, and frequency of data overlay; the implicit behavioral characteristics include frequency of use in high-data-consuming scenarios, proportion of time spent using high-frequency data-consuming apps, and density of fragmented internet access; the temporal variation characteristics include the data consumption fluctuation trends over the past 7 days, 30 days, and 90 days, and the behavioral activity change curve. The scene-related features include holiday traffic increase, outdoor scene traffic consumption ratio, and new APP usage traffic increase; The Pearson correlation coefficient combined with the random forest feature importance algorithm is used to calculate the correlation between each feature and the incremental traffic consumption. Core features with a correlation greater than a preset threshold are selected, redundant features and invalid noise features are removed, and the initial weight allocation is completed based on the feature contribution.

4. The method for determining potential traffic consumers based on AI according to claim 1, characterized in that, In step S3, the AI ​​model is an improved XGBoost multi-class prediction model. The model input is the user's core feature vector, and the output is the user's traffic consumption potential value in the range of 0-1. The sample dataset includes positive samples corresponding to the increase in traffic volume and negative samples corresponding to the decrease in traffic consumption. The model training process includes: dividing the sample dataset into training, validation, and test sets in a 7:2:1 ratio; initial model training using the training set; iterative optimization of model learning rate, tree depth, and iteration count hyperparameters using the validation set; validation of model accuracy, precision, and recall using the test set; and removal of noisy samples with errors exceeding a preset threshold until the overall model accuracy reaches the preset standard.

5. The method for determining potential traffic consumers based on AI according to claim 1, characterized in that, In step S4, the dynamic adaptive judgment threshold is generated based on the joint calculation of historical user identification accuracy, current marketing conversion rate, and real-time user traffic data, replacing the fixed judgment threshold. The user's traffic consumption potential value is compared with the dynamic adaptive judgment threshold. Users whose potential value is greater than or equal to the dynamic adaptive judgment threshold are judged as potential traffic consumption users.

6. The method for determining potential traffic consumers based on AI according to claim 1, characterized in that, In step S5, the potential level is divided into three levels: high potential, medium potential, and low potential, each corresponding to a different potential value range. Differentiated traffic operation strategies are matched for potential traffic consumers with different potential levels. Based on daily user consumption conversion data, the gradient descent algorithm is used to dynamically update the feature weights and model judgment thresholds to achieve adaptive iterative optimization of the model.

7. A system for identifying potential traffic consumers based on AI, characterized in that, The method for determining potential traffic consumers based on AI as described in any one of claims 1-6 includes a data acquisition and preprocessing module, an AI feature engineering module, a model training and optimization module, a user potential prediction module, a user hierarchical labeling module, and a dynamic iterative update module. The data acquisition and preprocessing module is used to collect multi-dimensional user raw data, and complete data cleaning, deduplication, normalization preprocessing, and output a standardized user dataset. The AI ​​feature engineering module is used to extract multi-dimensional traffic consumption features based on a standardized dataset, screen core effective features, and complete feature weight initialization. The model training and optimization module is used to build and train an improved XGBoost potential prediction AI model, complete model hyperparameter optimization and accuracy verification, and output a mature prediction model. The user potential prediction module is used to call the trained AI model to predict the traffic consumption potential value of the user to be identified, and to filter potential traffic consumption users by combining dynamic thresholds. The user stratification and labeling module is used to stratify the potential users after screening them by potential level and generate exclusive user profile tags. The dynamic iterative update module is used to dynamically iteratively optimize feature weights, model parameters, and judgment thresholds based on real-time operational conversion data, so as to achieve adaptive system updates.

Images