Parallel Data Domains for AI Model Drift Management
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Always-on applications with artificial intelligence models face challenges in handling varying user demand, maintaining high availability, and addressing data drift, which leads to significant downtime and inefficiencies.
Innovation Solution
The system employs a novel architecture that uses parallel data domains to track historical request data, detect concept and data drift, and automatically redeploy refit models, ensuring continuous service and preventing data loss.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the system uses a single data domain for model processing, then the system structure is simple, but data loss occurs and service availability decreases during model retraining
Solution Approach 1:
The patent divides the data domain into multiple parallel domains (first data domain, second data domain, third data domain) to store different types of data (inbound data, outbound data, training data). This segmentation allows the system to maintain service continuity by directing new requests to available domains while retraining models on other domains, thereby improving service availability without excessive complexity
Solution Approach 2:
The patent introduces a temporal dimension to data domain management by implementing domain rotation and switching mechanisms. The system transitions between different data domains over time (domain switching based on thresholds), adding a time-based dimension to the data architecture that enables continuous service while preparing and validating new models in parallel
2Measurement precision
If the system continuously monitors and detects data drift, then model performance is maintained, but system complexity and processing overhead increase
Solution Approach 1:
The patent implements preliminary data drift detection by continuously monitoring inbound data against the current model's expected input distribution before processing. The system calculates drift metrics (such as KL divergence or other statistical measures) in advance and maintains a buffer of detected drift events, enabling proactive model retraining decisions rather than reactive responses to performance degradation
Solution Approach 2:
The patent establishes a feedback loop where data drift detection results feed into automated model retraining triggers. The system continuously compares incoming data characteristics with training data distributions, and when drift thresholds are exceeded, the feedback mechanism automatically initiates model retraining processes, maintaining measurement precision through continuous monitoring while managing complexity through automated feedback-driven workflows
3Measurement precision
If the system retraines the model with detected outliers, then model accuracy improves, but service downtime increases
Solution Approach 1:
The patent prepares training data in advance by continuously collecting and buffering outbound data and detected outliers in the third data domain before they are needed for retraining. This preliminary data preparation ensures that when drift detection triggers a retraining event, the training data is already ready, significantly reducing the actual retraining time and service downtime
Solution Approach 2:
The patent implements domain switching that maintains continuous service by redirecting new requests to the first data domain (containing current valid model) while the second data domain undergoes retraining. The system continuously serves requests without interruption by switching between domains, ensuring that model retraining occurs in parallel with ongoing service rather than causing downtime
4Manufacturing precision
If the system stores and processes large amounts of historical data, then model retraining quality improves, but data storage and processing requirements increase
Solution Approach 1:
The patent applies local quality by storing different types of data with different retention requirements in separate data domains. The first data domain stores current valid inbound data, the second stores outbound data for retraining, and the third stores training data and outliers. This localized organization allows the system to retain only the specific portions of historical data needed for retraining (outliers and representative samples) rather than storing all historical data, improving retraining quality while managing storage efficiently
Solution Approach 2:
The patent implements a data lifecycle management system where data is discarded from active domains after serving its purpose and recovered when needed for retraining. Outliers and outbound data are retained in the third data domain for future retraining cycles, while older or less relevant data is discarded. This selective retention and recovery strategy ensures high-quality retraining data is available when needed while minimizing overall storage requirements
Data Source
AI summary
Systems and methods for a novel architecture to support always-on applications and/or models suffering drift in their results. The system may comprise one or more servers that are configured to track the historical behavior of incoming request data for a model and/or redirect the request as needed using parallel data domains. The one or more servers may maintain and update a catalog of potential data domains that partitions the historically received data. One partition may comprise data output from a current model. Another partition may comprise detected outliers in the data. In the case of drift, outliers, and/or anomalies in the incoming data, the system may return an error signal that causes data to be duplicated into a new data domain.