Automated Feature Generation via Formula Semantification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing approaches to automated feature generation in machine learning models are resource-intensive and often ineffective for smaller datasets, as they rely on brute force methods or require massive data, and are not suitable for diverse data types.
Innovation Solution
The method involves mapping existing features to mathematical formulas from a knowledge base, using natural language processing to generate new features by applying transformations based on matched formulas, allowing for efficient feature generation across various datasets.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If brute force approaches are used for automated feature generation, then feature generation can be achieved, but resource requirements (runtime, memory) increase significantly
Solution Approach 1:
The patent introduces a knowledge base as an intermediary that stores pre-computed feature transformations and relationships. Instead of brute-forcing all possible transformations, the system queries this knowledge base for relevant features based on data type and domain context, significantly reducing computational resources required while maintaining effective feature generation.
Solution Approach 2:
The knowledge base is pre-populated with domain-specific feature transformations and relationships before actual feature generation is needed. This preliminary action allows the system to retrieve pre-validated features directly during operation, avoiding the need to compute all possible transformations from scratch and reducing runtime and memory requirements.
2Productivity
If brute force approaches are used for automated feature generation, then feature generation can be achieved, but runtime increases significantly
Solution Approach 1:
The knowledge base serves as a pre-computed intermediary that stores valid feature transformations. During runtime, the system queries this pre-organized knowledge base rather than performing exhaustive searches, dramatically reducing the time required to generate features while maintaining comprehensive coverage of relevant transformations.
Solution Approach 2:
By pre-computing and storing feature transformations in the knowledge base before actual use, the system eliminates the need for time-consuming brute force computation during runtime. The preliminary action of pre-computation shifts the computational burden to a time when data is static, allowing rapid feature generation when data needs processing.
3Productivity
If existing approaches are used for automated feature generation, then feature generation is possible, but they are only effective for large datasets
Solution Approach 1:
The knowledge base acts as a domain-knowledge intermediary that encodes transformation rules and feature relationships independently of dataset size. This allows the same knowledge-based approach to effectively generate features for both small and large datasets, as the knowledge base provides domain-appropriate transformations regardless of the volume of input data.
Solution Approach 2:
The system adapts its behavior based on dataset characteristics by querying the knowledge base for appropriately scaled features. For small datasets, the system selects features that require less computational overhead, while for large datasets it can utilize more complex transformations, all guided by the domain knowledge in the knowledge base rather than brute-force exploration.
Data Source
AI summary
Embodiments for automated feature engineering are provided. Data associated with a machine learning model is received. The received data is mapped to at least one description associated with the data. A feature for the machine learning model is generated based on a formula within a corpus. The formula is associated with the at least one description.