Recursive Scoring Function for Multi-Value Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current decision support systems face challenges in effectively scoring and ranking data for multi-value classification problems, particularly in high-dimensional spaces, leading to suboptimal performance and incorrect use of distance to the margin as a score, which limits their ability to accurately predict outcomes in fields like finance and medicine.
Innovation Solution
A method is developed that builds a recursive scoring function which converges to an adaptive approximation of the optimal ROC curve, using resampling techniques and nonparametric approaches to generate interpretable scoring rules, allowing for visualization as an oriented binary tree and enabling the quantification of impact on each attribute, thus addressing the limitations of existing machine-learning algorithms.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional statistical methods (logistic regression, linear discriminant analysis) are used for scoring, then manual data preprocessing is required and the approach faces the curse of dimensionality, but these methods are simpler to implement and interpret
Solution Approach 1:
The patent extracts and eliminates the need for manual data preprocessing steps (modelling distribution, truncation of information) by directly applying machine learning algorithms to raw data. The scoring function is built recursively from the data itself without requiring preliminary statistical assumptions or manual feature engineering, thus taking out the problematic preprocessing stage entirely.
Solution Approach 2:
The patent replaces the mechanical statistical preprocessing system with an automated machine learning system. Instead of manually modelling distributions and truncating information, the system uses recursive algorithmic processing to directly learn scoring functions from data, substituting manual mechanical operations with automated computational processes.
2Adaptability or versatility
If binary classification methods are used, then the approach is simpler, but they incorrectly use the distance to the margin as score and cannot solve multi-value classification problems
Solution Approach 1:
The patent transforms the static binary classification approach into a dynamic recursive scoring system. Instead of a fixed decision boundary with margin distance, the system dynamically builds scoring functions through recursive partitioning that adapts to the data structure, allowing the scoring mechanism to evolve from simple binary classification to complex multi-value classification while maintaining accuracy.
Solution Approach 2:
The patent moves from the single-dimension binary classification (margin distance) to multi-dimensional scoring by recursively partitioning the feature space. This creates multiple dimensions of discrimination through hierarchical splitting, enabling the system to handle multi-value classification problems while providing nuanced scoring that goes beyond simple margin distance.
3Adaptability or versatility
If the number of dimensions is increased to handle more features, then the classification capability is improved, but the exponential increase in data volume causes the curse of dimensionality
Solution Approach 1:
The patent applies segmentation by recursively dividing the high-dimensional feature space into smaller, manageable partitions. Each recursive split divides the data volume exponentially less than the full space, making the problem tractable. This segmentation approach allows the system to handle high-dimensional data by breaking it down into a hierarchy of lower-dimensional subspaces that can be processed efficiently.
Solution Approach 2:
The patent transforms the high-dimensional curse into a manageable hierarchical structure by adding the dimension of recursion depth. Instead of processing all dimensions simultaneously in a flat structure, the system processes dimensions hierarchically across multiple recursive levels, effectively converting the exponential complexity of high-dimensional space into a manageable sequence of lower-dimensional processing steps.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The present invention is related to a method for providing with a score an object represented by a target item from a multidimensional space, said score being representative of the probability that the object satisfies a given binary property, said method being perform ed in a computer environment comprising processing means connected to memory means, comprising the following steps: (a) providing and storing in the memory means data representative of a set of sample items from said multidimensional space, each sample item being associated with a multi-value label representative of the probability that an object represented by said sample item satisfies the given binary property; (b) storing in the memory means data representative of the target item; (c) with said processing means, generating and storing in said memory means data representative of an oriented binary tree with a given height, said binary tree being defined by a plurality of nodes having index values in relation with their position in the binary tree, said nodes comprising a root node, intermediate nodes and terminal nodes, and each node being associated with a subspace of the multidimensional space according to the following rules: - if the node is a root node, the associated subspace is the whole multidimensional subspace; - if the node i s a non-terminal node, its associated subspace is split into two disjoint subspaces ε + and ε - respectively associated with a first child node and a second child node of said non-terminal node, said splitting being performed by the processing unit according to a classifier minimizing the number of sample items associated with a first given label value to belong to the subspace ε + , and/or the number of sample items associated with a second given label value to belong to the subspace ε - , said second label value being representative of a higher probability that an object represented by said sample item satisfies the given binary property than said first given label value; (d) with said processing means, associating with at least each terminal node of the binary tree a score depending on its index value in the binary tree, and storing the score value in the memory means; (e) with said processing means, identifying a target node, the target node being a terminal node whose associated subspace comprises the target item; and (f) outputting the score value associated with said target node. The invention is also related to a decision-support system.