Post-modeling Visualization for ML Feature Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing post-modeling visualization methods fail to provide detailed insights into the model's effect on specific portions of data, relying solely on numerical factors like feature importance and ROC curves.
Innovation Solution
A method that selects top N important features, discretizes continuous features, generates combination values, predicts target values, groups predictions, fits distributions, clusters distributions, and visualizes results using heat maps and distribution curves.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If traditional post-modeling visualization methods are used (relying on feature importance, ROC curves), then the overall model evaluation is provided, but detailed insights into model performance on specific data portions are not provided
Solution Approach 1:
The patent segments the training data into multiple data blocks based on feature value combinations. Each data block is visualized separately with its own distribution curve, allowing detailed inspection of model performance on specific data portions rather than providing only an aggregate view. This segmentation enables users to identify which specific data segments contribute to accurate or inaccurate predictions.
Solution Approach 2:
The patent adds a new visualization dimension by displaying distribution curves alongside tabular data. The x-axis represents prediction values, the y-axis shows density, and different colors represent different data blocks. This dimensional enhancement transforms traditional 1D or 2D visualizations into a multi-dimensional representation that simultaneously shows prediction distribution, data block membership, and feature relationships.
2Device complexity
If the number of distributions is reduced through clustering, then the visualization becomes more manageable, but the precision of individual data block representation may be compromised
Solution Approach 1:
The patent applies clustering algorithms to merge similar distributions into clusters, reducing the total number of distributions displayed. Each cluster is represented by a refitted distribution that captures the combined characteristics of member distributions. This merging reduces visualization complexity while preserving essential patterns through the refitted distributions.
Solution Approach 2:
For each cluster, the patent creates a refitted distribution that serves as a representative copy or approximation of the member distributions. These refitted distributions are displayed in the visualization, providing a simplified yet accurate representation of the underlying data block characteristics without requiring display of every individual distribution.
3Ease of manufacture
If continuous features are discretized into categories, then the combination values can be generated effectively, but information loss occurs during discretization
Solution Approach 1:
The patent changes the parameter representation of continuous features by discretizing them into categorical bins. This parameter transformation enables the generation of combination values that can effectively segment the data. The discretization process converts continuous parameter spaces into discrete categories, making it feasible to create meaningful data blocks for visualization while managing computational complexity.
Data Source
AI summary
In an approach, a processor selects a top N features for a machine learning (ML) model; discretizes values of each continuous feature of the top N features; generates a set of combination values that each represent a unique combination of feature values in for a data record; predicts, using the ML model, a target value for each record generating predicted target values; groups the predicted target values based on the combination value for each respective record; fits a distribution for each grouping of the predicted target values associated with a respective combination value generating a set of distributions; clusters and refits the set of distributions using a clustering algorithm resulting in a set of clusters and a refitted distribution for each cluster of the set of clusters; and outputs a visualization of the refitted distribution for each cluster as a distribution curve on a graph along with the associated records.