Online User Classification Model Management
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current online user classification methods are inefficient in accurately targeting advertisements, as they rely on deterministic classification, which limits the ability to classify users with limited past behavior, and require manual re-training of models, which is infrequent and not adaptive to real-time data.
Innovation Solution
The implementation of a system that automatically trains and updates classification models using stochastic gradient descent, allowing for real-time classification and distribution across a distributed computing system, enabling classification of users based on their characteristics and behavior, even with limited past activity, and facilitating ongoing model training with impressions received during operation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deterministic classification is used for user classification, then the classification process is simple and fast, but the accuracy is limited and cannot classify users with limited past behavior
Solution Approach 1:
The patent transitions from deterministic classification (binary yes/no based on fixed rules) to probabilistic classification using machine learning models that output likelihood scores. This parameter change enables the system to handle users with limited past behavior by providing continuous probability values rather than forced binary decisions, thereby improving classification accuracy without requiring complete user history.
Solution Approach 2:
The patent introduces machine learning models as intermediaries between raw user data and classification results. These models process and interpret user behavior patterns, converting complex behavioral data into meaningful classification probabilities. This intermediary layer enables accurate classification of users with limited past behavior while managing system complexity through modular model architecture.
2Adaptability or versatility
If manual re-training of classification models is performed, then the model can be updated with new data, but the process is infrequent and not adaptive to real-time data
Solution Approach 1:
The patent implements continuous learning where the classification model is trained incrementally on new user data as it arrives, rather than through periodic batch re-training. This continuous training process ensures the model adapts to changing user behaviors in real-time, maintaining high adaptability while efficiently utilizing computational resources through incremental updates rather than complete re-training cycles.
Solution Approach 2:
The system incorporates feedback mechanisms where classification results and user responses are fed back into the training process. The model continuously learns from new labeled data generated during operation, adjusting its parameters to improve future classifications. This feedback loop enables real-time adaptation without requiring manual intervention or stopping the classification service.
3Measurement precision
If classification models are re-trained on the entire dataset, then the model achieves maximum accuracy, but computational resources are excessively consumed
Solution Approach 1:
The patent applies incremental learning where the model is trained on a subset of new data (partial action) rather than the entire historical dataset. This approach maintains model accuracy by continuously incorporating recent relevant patterns while avoiding the excessive computational cost of periodic full-dataset re-training. The system processes only the necessary portion of data needed to adapt to recent changes in user behavior.
4Ease of operation
If deterministic classification rules are applied, then the system is easy to operate and interpret, but it cannot handle users with limited past behavior effectively
Solution Approach 1:
The patent segments the classification approach into multiple components: feature extraction modules that process different user attributes independently, multiple classification models for different user types, and a scoring system that aggregates results. This segmentation allows the system to handle users with limited behavior data by using alternative features and models, maintaining ease of operation through modular design while improving accuracy for edge cases.
Data Source
AI summary
Methods and apparatus for managing models for classification of online users are disclosed. An example method includes determining a segment associated with a received impression, determining whether the received impression includes an indication that the received impression is the result of a conversion associated with the segment, determining a current ratio of converted to non-converted impressions associated with a classification model associated with the segment, determining a difference between the current ratio and a target ratio, and storing an indication of the received impression for use in at least one of training or updating a classification model associated with the segment when: (a) the difference does not meet a threshold or (b) counting the received impression in the ratio would reduce the difference.