Binary Linear Classification via Logistic Regression
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine-learning classifier algorithms for binary classification are sensitive to the number of features, require iterative training, and struggle to effectively utilize metadata, making them inefficient for large datasets and streaming systems.
Innovation Solution
A discriminative binary text classification model that is scalable, computationally efficient, and self-tuning, using filtered logistic regression and incorporating both word counts and metadata, which improves performance with added features and employs an array of classifiers to sort documents based on topics.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If existing machine-learning classifier algorithms are used for binary classification, then classification can be performed, but the system becomes sensitive to the number of features and requires iterative training
Solution Approach 1:
The patent changes the mathematical parameters of the classification model by using logistic regression with a simplified decision boundary approach. Instead of complex iterative algorithms, the system uses a linear combination of features weighted by coefficients, transformed through a sigmoid function. This parameter transformation enables non-iterative training while maintaining classification accuracy and reducing sensitivity to feature dimensionality.
2Quantity of substance
If existing classifier algorithms process large datasets, then classification can be performed, but computation time increases significantly
Solution Approach 1:
The patent replaces complex mechanical iterative optimization processes with a direct mathematical computation system. Instead of using gradient descent or other iterative algorithms that require multiple passes through the data, the system computes classification results through a single pass using logistic regression equations. This substitution of computational mechanics enables processing of large datasets with reduced time complexity.
3Speed
If existing algorithms process streaming data, then real-time classification can be achieved, but the algorithms struggle to effectively utilize metadata
Solution Approach 1:
The patent creates a universal classification framework that handles multiple data types uniformly. The logistic regression model processes textual features, metadata fields, and other data attributes through the same mathematical formulation. This multi-functional approach allows the system to effectively utilize metadata alongside traditional features while maintaining fast processing speeds suitable for streaming data.
4Reliability
If more features are added to improve classification accuracy, then performance improves, but the system becomes more sensitive to feature dimensionality
Solution Approach 1:
The patent transforms the classification problem by introducing a new mathematical dimension through the logistic sigmoid function. This transformation maps the linear feature space into a non-linear decision space, where the relationship between features and class labels is captured through the sigmoid transformation. This dimensional transformation reduces sensitivity to feature dimensionality while maintaining or improving accuracy with additional features.
Data Source
AI summary
In an example, there is disclosed a computing apparatus for performing binary classification on a corpus of objects, having: one or more logic elements providing a classification engine to: access a training set comprising a plurality of binary classified objects, each object including one or more features, and each object designated as belonging to the class or not belonging to the class; build a model comprising a probability that each of the one or more features implies that an inspected object belongs to the class; access a particular object from the corpus; analyze the particular object to identify one or more identified features, the identified features selected form the one or more features; assign the particular object a composite score comprising a sum of the identified features; and designate the particular object as belonging to the class if and only if the composite score is above a threshold.