Exploring Class Overlap in Classification Challenges: Introducing the R Package
useR! 2024, Salzburg, Austria
Priyanga Dilini Talagala
July 9, 2024
Data Quality Problems
Class Overlapping Problem
Class overlap occurs when instances of more than one class share a common region in the data space and are not clearly separable
Class Overlapping Problem
Class overlap occurs when instances of more than one class share a common region in the data space and are not clearly separable
This overlap can happen due to:
Inherent Similarity: Natural similarity between classes
Class Overlapping Problem
Class overlap occurs when instances of more than one class share a common region in the data space and are not clearly separable
This overlap can happen due to:
Inherent Similarity: Natural similarity between classes
Noise: Variability or errors in data collection
Class Overlapping Problem
Class overlap occurs when instances of more than one class share a common region in the data space and are not clearly separable
This overlap can happen due to:
Inherent Similarity: Natural similarity between classes
Noise: Variability or errors in data collection
Feature Representation: Insufficient or inadequate features to separate classes
Implications of Class Overlap
Classifiers struggle to correctly classify instances due to overlapping regions
Implications of Class Overlap
Classifiers struggle to correctly classify instances due to overlapping regions
Higher error rates occur in areas where classes overlap, leading to more instances being misclassified
Implications of Class Overlap
Classifiers struggle to correctly classify instances due to overlapping regions
Higher error rates occur in areas where classes overlap, leading to more instances being misclassified
If the problem of class overlap is not addressed, models may become overly complex, leading to overfitting issues where the model performs well on training data but poorly on unseen data
Types of Class Overlapping Problems
clap R Package
Detecting Class OverLAPping Regions in Multidimensional Data
install.packages('clap')
devtools::install_github("pridiltal/clap")
clap Framework
Normalize the columns of the data. (median and IQR)
This prevents variables with large variances having disproportional influence on Euclidean distances.
clap Framework
Leader Algorithm (Hartingan, 1975)
Calculate the nearest neighbor distances
clap Framework
Leader Algorithm (Hartingan, 1975)
Calculate the nearest neighbor distances
Performs clustering using a radius based on the maximum nearest neighbor distance
Case Study 1: Biopsy Data of Breast Cancer Patients
clump thickness
uniformity of cell size
uniformity of cell shape
marginal adhesion
single epithelial cell size
bare nuclei (16 values are missing)
bland chromatin
normal nucleoli
mitoses
Classification Task
Benign tumor - generally do not invade and spread
Malignant tumor cells - more likely to spread to other areas of the body
Case Study 2: Detecting Class Overlapping Instances of Image databases
Aim: detect early stages of pneumonia through chest x-ray images
Case Study 2: Detecting Class Overlapping Instances of Image databases
Aim: detect early stages of pneumonia through chest x-ray images
Inaccurate analysis of the x-ray results may result in an improper diagnosis and decision making, resulting in a costly mistake that could otherwise save lives
Case Study 2: Detecting Class Overlapping Instances of Image databases
Aim: detect early stages of pneumonia through chest x-ray images
Inaccurate analysis of the x-ray results may result in an improper diagnosis and decision making, resulting in a costly mistake that could otherwise save lives
Linearized chest x-rays images
Images of Class “Normal”
Images of Class “Pneumonia”
What Next?
The Leader Algorithm determined cluster size using the maximum nearest neighbor distance. Further analysis is needed to identify the optimal cluster size.
What Next?
The Leader Algorithm determined cluster size using the maximum nearest neighbor distance. Further analysis is needed to identify the optimal cluster size.
Euclidean distance was used in the Leader Algorithm to determine cluster size. Additional experiments are required to find faster distance calculation methods.
What Next?
The Leader Algorithm determined cluster size using the maximum nearest neighbor distance. Further analysis is needed to identify the optimal cluster size.
Euclidean distance was used in the Leader Algorithm to determine cluster size. Additional experiments are required to find faster distance calculation methods.
Further analysis is needed to identify classification methods that are robust to class overlapping instances.
What Next?
The Leader Algorithm determined cluster size using the maximum nearest neighbor distance. Further analysis is needed to identify the optimal cluster size.
Euclidean distance was used in the Leader Algorithm to determine cluster size. Additional experiments are required to find faster distance calculation methods.
Further analysis is needed to identify classification methods that are robust to class overlapping instances.
Further analysis is required to assess the overlapping status of unseen test data in the high-dimensional feature space.
Thank you
This work was supported in part by the RETINA research lab, funded by the OWSD, a program unit of the United Nations Educational, Scientific, and Cultural Organization (UNESCO).
