Detuned Decision Tree Tuning for Accurate Entity Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing decision trees for entity selection are prone to overfitting or underfitting, and existing pruning algorithms are insensitive to feature uniqueness and context, leading to inaccurate identification of entities that may benefit from a target feature.
Innovation Solution
A detuned decision tree generation mechanism that selects a subset of trainable features based on feature metadata, sets threshold quantities for leaf nodes, and adjusts hyperparameters to ensure the tree predicts a target audience size within specified limits, generating queries to identify entities lacking the target feature.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a decision tree is trained to accurately predict target feature, then prediction accuracy is improved, but the model becomes prone to overfitting
Solution Approach 1:
The patent changes the parameters of the decision tree by introducing detuning factors that modify the splitting criteria and leaf node selection. Instead of purely optimizing for prediction accuracy, the system adjusts parameters like minimum samples per leaf, maximum tree depth, and splitting thresholds to prevent overfitting while maintaining acceptable accuracy.
Solution Approach 2:
The patent applies partial action by intentionally not fully optimizing the decision tree for prediction accuracy. It uses pruning techniques and sets constraints that prevent the model from learning all patterns in the training data, thereby sacrificing some accuracy to improve generalization and avoid overfitting.
2Device complexity
If existing pruning algorithms are applied to decision trees, then model complexity is reduced, but the algorithms are insensitive to feature uniqueness and context leading to inaccurate entity identification
Solution Approach 1:
The patent applies local quality by treating different features and contexts uniquely during the pruning process. Instead of applying uniform pruning criteria, it considers feature metadata, data types, and contextual relationships to selectively prune or retain specific branches based on their unique characteristics and contribution to accurate entity identification.
Solution Approach 2:
The patent performs preliminary actions by analyzing feature metadata and contextual relationships before applying pruning. It pre-identifies important features and contextual patterns that should be preserved, then uses this information to guide the pruning process, ensuring that critical information is not lost while still reducing model complexity.
3Adaptability or versatility
If a decision tree includes all features to capture comprehensive patterns, then prediction coverage is improved, but the model becomes overly complex and computationally expensive
Solution Approach 1:
The patent extracts only the necessary features and patterns needed for effective entity selection. It uses feature metadata and contextual analysis to identify and retain only the most relevant features, removing redundant or less important ones, thereby maintaining prediction coverage while reducing model complexity and computational cost.
Solution Approach 2:
The patent segments the feature set into different groups based on their importance, data type, and contextual relevance. It processes and evaluates features in segments rather than all at once, allowing for more manageable computation while ensuring comprehensive coverage of important patterns through systematic feature selection.
Data Source
AI summary
A subset of trainable features from a set of features is selected based at least in part on feature metadata. A trained decision tree is generated to predict which entities of a plurality of entities have a target feature. A corresponding quantity of entities of the plurality of entities that do not have the target feature is determined for each leaf node in the trained decision tree. A subset of the plurality of leaf nodes are selected wherein each leaf node in the subset of leaf nodes has a corresponding quantity of entities in between a minimum threshold quantity and a maximum threshold quantity. An aggregate quantity of entities is determined based on a sum of the corresponding quantity of entities of each leaf node in the subset of leaf nodes. The process is repeated until a suitable trained decision tree is generated.