Nominal Feature Transformation via Outcome Likelihood Encoding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Nominal features in machine learning models are difficult to utilize due to their non-numeric nature, leading to increased data dimensionality and storage requirements when transformed into Boolean features, which can significantly impact model size and processing efficiency.
Innovation Solution
Transforming nominal features into numeric features representing the likelihood or probability of an outcome, reducing data dimensionality and enabling more efficient model generation and prediction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If nominal features are transformed into Boolean features to represent each potential value, then the nominal feature can be used in machine learned models, but the size of the input data significantly increases
Solution Approach 1:
The patent transforms nominal features by changing the parameter representation from Boolean (0/1) to numeric values representing the count or frequency of each nominal value in the training data. This parameter change allows the model to capture distribution information while using fewer features than the traditional one-hot encoding approach.
Solution Approach 2:
Instead of expanding nominal features into multiple Boolean dimensions (one-hot encoding), the patent compresses them into a single numeric dimension that encodes frequency or count information. This dimensional transformation reduces feature space while preserving essential information about the nominal feature distribution.
2Ease of operation
If nominal features are transformed into Boolean features, then the model can process the data, but storage and processing requirements increase
Solution Approach 1:
The patent changes the parameter type from Boolean to numeric (count/frequency), which maintains mathematical operability for machine learning models while significantly reducing the number of features. This allows standard mathematical operations to continue working effectively with fewer dimensions.
Solution Approach 2:
The patent merges multiple Boolean features that would represent each nominal value into a single numeric feature that captures the aggregate count or frequency information. This consolidation reduces the number of features from N (number of nominal values) to 1, simplifying storage and processing.
3Loss of information
If multiple Boolean features are created for each nominal feature, then each potential value is represented, but the number of features increases significantly
Solution Approach 1:
The patent transforms the representation parameter from presence/absence (Boolean) to frequency/count (numeric), enabling a single feature to encode information about all nominal values through its numeric value, thereby reducing feature count while maintaining information density.
Data Source
AI summary
Embodiments of the present invention relate to transforming a nominal feature to a numeric feature that indicates a likelihood or probability of a particular outcome. Numeric features are determined that indicate a likelihood of an outcome given the value of the collected data (nominal values). Such numeric features are used to represent the corresponding nominal features for use in generating a machine learned model. As such, a nominal feature initially captured in a data set is transformed or converted to a numeric feature that represents a likelihood of a corresponding outcome as opposed to a Boolean value. Upon transforming nominal values to numeric values based on the likelihood of outcome, the numeric values can be used to generate a machine learned model that is used to predict future outcomes.