THE IMPACT OF THE ORDER OF PRESENTING SAMPLES FROM DIFFERENT DIAGNOSTIC CLASSES ON THE TRAINING AND EVALUATION OF CLASSIFIERS

This article investigates the impact of the presentation order of samples from different diagnostic classes on the classification results of biomedical data. The objective is to determine whether classifier performance metrics change when the order of training records is altered, while keeping the composition of the training and test sets constant. The study utilizes the Wisconsin Diagnostic Breast Cancer (WDBC) dataset, which contains 569 samples characterized by 30 numerical features: 212 malignant cases and 357 benign cases. The training set consisted of 455 records, and the test set comprised 114 records (42 malignant and 72 benign). Three ordering schemes were compared: presenting malignant samples before benign ones (M – B), presenting benign samples before malignant ones (B – M), and random shuffling (using a random seed of 42). The feature values and class labels remained unchanged. Comparing these orderings while maintaining a fixed dataset composition allows for isolating the effect of the training sequence from the consequences of reallocating records between the training and test sets. Eighteen experimental runs were conducted within the WEKA environment using the following algorithms: J48, RandomForest, SMO, MultilayerPerceptron, HoeffdingTree, and SGD (with a hinge loss function). The configuration settings for each algorithm remained consistent across all ordering schemes. The results were evaluated based on accuracy, confusion matrices, and the recall (sensitivity) for detecting malignant cases. The most significant differences were observed for the Multilayer Perceptron: accuracy was 92.98% for the M – B order, 97.37% for the B – M order, and 93.86% for the mixed order. The difference between the highest and lowest values was 4.39 percentage points, corresponding to five correct predictions. For RandomForest and SGD, the variation in accuracy was 0.88 percentage points, representing a single test sample. For J48, SMO, and HoeffdingTree, the accuracy and confusion matrices remained identical across all three orderings. It is demonstrated that identical overall accuracy can be accompanied by different distributions of errors across classes. For instance, RandomForest achieved an accuracy of 95.61% under both the B – M and mixed orderings; however, it misclassified two and three malignant samples as benign, respectively. These results indicate that the classifier implementations studied exhibit varying sensitivity to the input order of records. Random shuffling did not yield the highest accuracy for all models. These findings pertain to a single data split and a fixed set of parameters; they do not establish a statistical advantage for any specific ordering. Further validation should involve repeated runs using different random seed values and multiple data splits, alongside monitoring the actual sequence in which examples are processed within WEKA.

Authors

Institutions

Publication Details

Journal
Middle European Scientific Bulletin
Published
2026-11-13
DOI
https://doi.org/10.65237/2336-3630-2026-11(1)-5
Primary Topic
Artificial Intelligence in Healthcare
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

THE IMPACT OF THE ORDER OF PRESENTING SAMPLES FROM DIFFERENT DIAGNOSTIC CLASSES ON THE TRAINING AND EVALUATION OF CLASSIFIERS

Denys Honcharov
1 links
Middle European Scientific Bulletin
Artificial Intelligence in Healthcare
article

THE IMPACT OF THE ORDER OF PRESENTING SAMPLES FROM DIFFERENT DIAGNOSTIC CLASSES ON THE TRAINING AND EVALUATION OF CLASSIFIERS

Denys Honcharov
article en

Abstract

This article investigates the impact of the presentation order of samples from different diagnostic classes on the classification results of biomedical data. The objective is to determine whether classifier performance metrics change when the order of training records is altered, while keeping the composition of the training and test sets constant. The study utilizes the Wisconsin Diagnostic Breast Cancer (WDBC) dataset, which contains 569 samples characterized by 30 numerical features: 212 malignant cases and 357 benign cases. The training set consisted of 455 records, and the test set comprised 114 records (42 malignant and 72 benign). Three ordering schemes were compared: presenting malignant samples before benign ones (M – B), presenting benign samples before malignant ones (B – M), and random shuffling (using a random seed of 42). The feature values and class labels remained unchanged. Comparing these orderings while maintaining a fixed dataset composition allows for isolating the effect of the training sequence from the consequences of reallocating records between the training and test sets. Eighteen experimental runs were conducted within the WEKA environment using the following algorithms: J48, RandomForest, SMO, MultilayerPerceptron, HoeffdingTree, and SGD (with a hinge loss function). The configuration settings for each algorithm remained consistent across all ordering schemes. The results were evaluated based on accuracy, confusion matrices, and the recall (sensitivity) for detecting malignant cases. The most significant differences were observed for the Multilayer Perceptron: accuracy was 92.98% for the M – B order, 97.37% for the B – M order, and 93.86% for the mixed order. The difference between the highest and lowest values was 4.39 percentage points, corresponding to five correct predictions. For RandomForest and SGD, the variation in accuracy was 0.88 percentage points, representing a single test sample. For J48, SMO, and HoeffdingTree, the accuracy and confusion matrices remained identical across all three orderings. It is demonstrated that identical overall accuracy can be accompanied by different distributions of errors across classes. For instance, RandomForest achieved an accuracy of 95.61% under both the B – M and mixed orderings; however, it misclassified two and three malignant samples as benign, respectively. These results indicate that the classifier implementations studied exhibit varying sensitivity to the input order of records. Random shuffling did not yield the highest accuracy for all models. These findings pertain to a single data split and a fixed set of parameters; they do not establish a statistical advantage for any specific ordering. Further validation should involve repeated runs using different random seed values and multiple data splits, alongside monitoring the actual sequence in which examples are processed within WEKA.

Middle European Scientific BulletinVol. 1(11)
Petro Mohyla Black Sea National University (UA)
Artificial Intelligence in Healthcare
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon