Event-Based Feature Engineering for ML Training Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine learning algorithms face challenges in generating event-based feature vectors and training examples for accurate point-in-time predictions, as data scientists lack access to event-based data and must spend significant time writing complex code to manipulate data, leading to labor-intensive and inefficient model training and deployment.
Innovation Solution
A feature engineering system that ingests data from various sources, computes features in real-time, and provides a user-friendly interface for data scientists to define feature selection and configuration, enabling the generation of event-based feature vectors and training examples at arbitrary points in time, facilitating accurate and scalable model training and deployment.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If data scientists manually manipulate event-based data to generate feature vectors, then model accuracy can be improved, but the time and labor required increases significantly
Solution Approach 1:
The system pre-computes and stores feature values in a feature store before they are needed for model training or deployment. Event-based features are calculated and cached in advance, allowing data scientists to retrieve pre-computed features without manually manipulating raw event data, thus improving model accuracy while reducing the time required for feature engineering
Solution Approach 2:
A feature store acts as an intermediary layer between raw event data and machine learning models. The feature store stores pre-computed event-based features that can be directly consumed by models, eliminating the need for data scientists to write complex manipulation code while ensuring accurate feature vectors are provided to models
2Ease of manufacture
If complex code is written to manipulate event-based data, then feature vectors can be generated, but the complexity and labor intensity increases
Solution Approach 1:
The feature store automatically computes and stores event-based features without requiring data scientists to write manipulation code. The system self-services by ingesting event data, computing relevant features according to pre-defined schemas, and making them available for model training, thereby reducing both labor intensity and code complexity while maintaining feature generation capability
Solution Approach 2:
The feature store provides a universal interface for generating event-based features that can be used across multiple machine learning models and applications. Instead of writing separate manipulation code for each model, the feature store serves as a multi-functional system that can generate features for various models, reducing overall system complexity and labor requirements
3Measurement precision
If feature values are updated in real-time, then model predictions become more accurate, but the computational resources and system complexity increase
Solution Approach 1:
The system pre-computes feature values and stores them in the feature store before they are needed for predictions. By calculating features in advance and caching them, the system can provide up-to-date feature values without requiring complex real-time computation infrastructure, thus improving prediction accuracy while managing system complexity
Solution Approach 2:
The feature store serves as an intermediary that handles the complexity of feature computation and storage. It manages the updating and retrieval of feature values, providing a simplified interface to models while handling the computational complexity in the background, thereby enabling accurate predictions without directly increasing model system complexity
Data Source
AI summary
A method for generating machine learning training examples using data indicative of events associated with a plurality of entities. The method comprises receiving an indication of one or more selected entities of the plurality of entities, receiving information indicative of selecting one or more prediction times associated with each of the one or more selected entities, and receiving information indicative of selecting one or more label times associated with each of the one or more selected entities. Each of the one or more label times corresponds to at least one of the one or more prediction times, and the one or more label times occur after the corresponding one or more prediction times. Data associated with the one or more prediction times and the one or more label times is extracted from the data indicative of events associated with the plurality of entities. Training examples for use with a machine learning algorithm are generating using the data associated with the one or more prediction times and the one or more label times.