Auto Feature Preparation for Online Inferencing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing predictive analysis systems face performance bottlenecks due to inefficient data transformation processes, which often require more time than machine learning model evaluation, and existing solutions focus on optimizing prediction components rather than data preparation.
Innovation Solution
The method involves analyzing and optimizing data transformation operators during machine learning model training, creating multiple data preparation pipelines, and deploying them to a runtime system, including batch, streaming, and online pipelines to manage data refresh rates and latency requirements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of time
If data transformation is performed using traditional ETL processes, then data preparation can be completed, but latency requirements cannot be met and machine learning models cannot be used in real-time scenarios
Solution Approach 1:
The patent segments the data preparation process into multiple independent pipelines (batch pipeline, streaming pipeline, online pipeline) that can operate simultaneously with different latency characteristics. Each pipeline handles specific data transformation operations optimized for its latency requirements, allowing the system to meet real-time inferencing needs while maintaining comprehensive data preparation capabilities.
Solution Approach 2:
The system dynamically selects and activates appropriate pipelines based on the specific inferencing requirements. The framework automatically determines which pipelines to use for batch processing versus real-time processing, enabling flexible adaptation to different latency requirements without manual configuration of complex ETL processes.
2Productivity
If custom data transformation applications are designed and implemented manually, then data preparation optimization is achieved, but development costs and complexity increase significantly
Solution Approach 1:
The framework performs self-configuration and automatic pipeline generation based on the provided data transformation operations. Instead of requiring manual development of custom applications, the system automatically creates optimized pipelines, assigns them to appropriate pipeline types, and configures execution parameters, significantly reducing development effort while maintaining performance optimization.
Solution Approach 2:
The system automatically adjusts execution parameters such as data refresh rates, pipeline selection, and resource allocation based on the characteristics of data transformation operations. This automated parameter optimization eliminates the need for manual tuning and reduces development complexity while achieving performance gains.
3Productivity
If existing feature cache solutions are used, then performance is improved for specific applications, but adaptability to general solutions is limited
Solution Approach 1:
The framework provides a universal solution that can handle multiple types of data transformation operations and inferencing scenarios through a single unified architecture. The system supports batch, streaming, and online pipelines that can be combined in various configurations to meet different application requirements, making it adaptable to general solutions rather than application-specific implementations.
Data Source
AI summary
An approach is provided in which a method, system, and program product analyze, while training a machine learning model, a set of first data transformation operators in a first data preparation pipeline that generates a plurality of constructed features from a set of training data. The method, system, and program product create a plurality of second data preparation pipelines from the first data preparation pipeline, wherein the set of first data transformation operators are converted to a set of second data transformation operators and each assigned to one of the plurality of second data preparation pipelines. The method, system, and program product deploy the plurality of second data preparation pipelines to a runtime system.