KNN and Naive Bayes Weighted Polling for Dry Bean Classification

Source code supporting:

Morsidi, F. (2026). Hybrid machine learning merging KNN and Naive Bayes weighted polling for dry beans multiclass classification. ASM Science Journal, 21(2), 1-18. https://doi.org/10.32802/asmscj.2026.66

What this is

A Java implementation of a K-Nearest Neighbours classifier with four polling methods, evaluated on the Dry Bean dataset (13,611 samples, 7 classes, 8 morphological features) with leave-one-out testing and exhaustive feature-subset search.

Method	Polling rule
1	Normal: each neighbour contributes one vote
2	Weighted: closer neighbours contribute more
3	Weighted-Squared: closer neighbours contribute quadratically more
4	Naive Bayes style: votes normalised by class frequency

Methods 1-3 correspond to Table 3 of the paper.

Files

• src/gnnbeanfix/Classifier.java - classifier, polling methods, leave-one-out test, feature-subset search
• src/gnnbeanfix/BeanDataFileRead.java, ArrayListMath.java - data loading and helpers
• src/gnnbeanfix/GUI.java, HelpGUI.java, GraphCreator.java - Swing interface and parallel-coordinates chart (requires JFreeChart)
• src/gnnbeanfix/Main.java - launches the GUI; also holds helper methods for batch accuracy searches
• src/gnnbeanfix/Reproduce.java - reproduces the accuracy values reported in the paper
• TestBeanDataset1.txt - the dataset in the format the code reads

Classifier, BeanDataFileRead, ArrayListMath and Reproduce need only the standard Java library. The GUI and chart classes need JFreeChart on the classpath.

Reproducing the reported accuracies

From the project root (the folder containing TestBeanDataset1.txt):

javac -d bin src/gnnbeanfix/Classifier.java src/gnnbeanfix/BeanDataFileRead.java src/gnnbeanfix/ArrayListMath.java src/gnnbeanfix/Reproduce.java
java -cp bin gnnbeanfix.Reproduce

Setup: leave-one-out, k = 3, polling method 3. Expected output (each line takes roughly 20-60 seconds):

All 8 features         accuracy = 88.54%  (paper: 88.54%)
Without Perimeter      accuracy = 85.67%  (paper: 85.67%)
Without EquivDiameter  accuracy = 88.58%  (paper: 88.58%)

Feature order in the code: Area, Perimeter, MajorAxisLength, MinorAxisLength, ConvexArea, EquivDiameter, ShapeFactor1, ShapeFactor3.

Notes

• The reported accuracies were reproduced with OpenJDK 21.
• The code implements the vote-tally polling methods above. It does not train a local Naive Bayes model per test point or derive classifier weights by cross-validation.
• Figure 4 in the paper labels Perimeter as included in every iteration. Running this code, the 85.67% result arises when Perimeter is excluded.

Data

TestBeanDataset1.txt is derived from the Dry Bean dataset by Koklu and Ozkan (2020), Computers and Electronics in Agriculture, 174, 105507, https://doi.org/10.1016/j.compag.2020.105507, obtained from the UCI Machine Learning Repository. The original dataset is licensed under Creative Commons Attribution 4.0 International (CC BY 4.0), https://creativecommons.org/licenses/by/4.0/.

Changes made: the file keeps all 13,611 samples but only 8 of the original 16 features (Area, Perimeter, MajorAxisLength, MinorAxisLength, ConvexArea, EquivDiameter, ShapeFactor1, ShapeFactor3) plus the class label, stored as space-separated text.

The data file remains under CC BY 4.0. The licence below applies to the source code only.

Licence

The source code is released under the MIT License (see LICENSE). Copyright (c) 2026 Farid Morsidi.