Predicting Binary Outcomes Using Partial Mutual Information
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Database systems face challenges in efficiently processing large volumes of data to build predictive models for binary outcomes due to the vast number of samples and features, making it difficult to determine predictive value effectively.
Innovation Solution
A method is introduced to select a limited set of predictive terms by creating positive and negative ordered pools based on partial mutual information, reducing the candidate pool through strength and diversity analysis, and constructing a predictive model using these terms.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If all samples and features are processed to build a predictive model, then model completeness is improved, but computational complexity increases
Solution Approach 1:
The patent extracts only the most predictive terms from the large set of candidate terms by calculating partial mutual information scores. Terms with scores above a threshold are selected for the final model, removing irrelevant or redundant features before model construction. This extraction process reduces the input data size while preserving the most valuable predictive signals.
Solution Approach 2:
The patent segments the term selection process into distinct stages: initial term generation, partial mutual information calculation, threshold-based filtering, and final model construction. This segmentation allows the system to handle the large dataset in manageable steps, processing only the most relevant terms at each stage rather than all terms simultaneously.
2Reliability
If a large number of terms are used in the predictive model, then predictive coverage is improved, but processing time increases
Solution Approach 1:
The patent applies partial action by selecting only a subset of terms that meet the partial mutual information threshold, rather than using all possible terms. This partial selection achieves sufficient predictive coverage for the model while significantly reducing the number of terms that need to be processed, thereby decreasing processing time without sacrificing essential predictive capability.
Solution Approach 2:
The patent changes the parameter of term selection from including all terms to including only terms with partial mutual information scores above a specific threshold. This parameter change filters out low-value terms before model construction, reducing the computational burden while maintaining the model's ability to capture important predictive relationships.
3Reliability
If redundant terms are included in the model, then feature coverage is improved, but model efficiency decreases
Solution Approach 1:
The patent introduces partial mutual information as an intermediary metric to evaluate and filter candidate terms. This intermediary measure acts as a gatekeeper, allowing only terms with sufficient predictive value (above the threshold) to proceed to the final model. This intermediary filtering step eliminates redundant terms while preserving feature coverage provided by high-value terms.
Solution Approach 2:
The patent extracts and removes redundant terms from the candidate pool by applying the partial mutual information threshold. Terms that do not meet the threshold are extracted and discarded, while only the essential predictive terms are retained for model construction. This extraction process ensures the final model includes diverse features without redundant information.
Data Source
AI summary
A binary outcome of an activity is predicted based on samples of the activity. The activity is characterized by features that can take on any of a set of mutually exclusive levels. An initial candidate pool of terms is selected. The terms are feature levels or combinations of feature levels. The candidate terms are used to create two ordered pools of terms, one including terms that are positively predictive and another including terms that are negatively predictive. The terms in each pool are ordered by strength of predictiveness and diversity of predictiveness relative to terms higher in the order. A final set of terms is selected by combining terms from these two pools.