Source-linked AI summary
Classification Based on Association Rules Algorithm for Breast Cancer
Ali Alsalama, Ahmed Kubba, Ghaith Jamjoum, Zaher Al Aghbari
TL;DR
Early breast cancer detection motivates the paper’s classification approach. It presents a weighted association-rule implementation with generation, pruning, and prediction stages, and reports support-level patterns alongside correct prediction of a test sample’s class.
Problem
Early cancer diagnosis is presented as important because it may significantly reduce mortality.
Source: B. Breast Cancer · p. 2Method
The implementation generates association rules, prunes them into major and minor groups, and predicts test-instance classes using the pruned rules.
Source: D. Contributions · p. 2, B. Rule Pruning · p. 4, C. Rule Prediction · p. 4Results
Accuracy and F1 are reported as highest at 20% support in Categories 1 and 3, while support around 10% is reported as ideal for Category 2.
Source: C. Experimental Results · pp. 5–6Takeaways & Limitations
The paper reports that its prediction algorithm correctly classified the tested sample, supporting the feasibility of this breast-cancer association-rules classifier.
Source: CONCLUSION · p. 6
Abstract
from arXiv · showhide
Breast cancer is a significant contributor to female mortality across the world, displaying one of the highest oc currence rates among the various cancer types. In response to the need for early breast cancer detection, researchers have increasingly turned to association rule-based classification as a favored method. Association Rule mining is a data mining approach which offers the benefit of yielding results that are readily understandable for medical professionals. This paper introduces a novel association rule-based data mining technique for breast cancer classification based on a weighted classification approach. This implementation employs three core algorithms: Rule Generation, Rule Pruning, and Rule Prediction. Rule Generation identifies frequent itemsets and creates association rules. Rule Pruning eliminates rules using specific criteria and separates them into major and minor groups based on their influence on training data. Rule Prediction applies the pruned rules to classify test data. The final prediction algorithm was tested on several testing samples to show the feasibility and performance of the approach.
I. INTRODUCTION
Breast cancer’s prevalence and mortality motivate classification efforts [2], and association classification (AC) integrates association rules into classification.
I. INTRODUCTION
- Association classification (AC) integrates association-rule mining into classification to uncover patterns and potentially enhance accuracy.The approach connects attributes with class labels through rules such as co-occurring attributes predicting a class.Source: §I · p. 1
A. Association Classification
- Association methods uncover relationships among items, while classification predicts an instance’s class after training on a dataset.The paper describes AC as a fusion of these two functions.Source: A. Association Classification · p. 1, §I · p. 1
- Support is defined by rule support count over total transactions, while confidence divides support count by the rule antecedent’s occurrence count.The passages define support count and antecedent occurrence using rows in the training dataset.Source: A. Association Classification · pp. 1–2
- A class association rule has an itemset as its antecedent and a class as its consequent, with support count tracking matching rows assigned to that class.The actual occurrence counts rows matching the itemset, regardless of class.Source: A. Association Classification · pp. 1–2
- Frequent itemset rules meet the minimum-support threshold, and candidate rules are also checked against minimum-support and minimum-confidence conditions.The paper introduces support and confidence as measures used to generate association rules.Source: A. Association Classification · p. 2
B. Breast Cancer
The paper frames breast cancer as a reason to prioritize early diagnosis, and describes recurrence prediction as a data-mining application using patient indicators.
B. Breast Cancer
- Early diagnosis is presented as a major medical concern because it may significantly reduce mortality for diseases such as cancer.The paper distinguishes benign and malignant cancers.Source: B. Breast Cancer · p. 2
- Breast cancer is described as a leading cancer among women globally and as the second leading cause of cancer-related deaths in women.The passage also states that about 10% of women worldwide experience it at some point in life.Source: B. Breast Cancer · p. 2
- The paper describes recurrence prediction as a complex task and notes that data mining can use indicators derived from statistical results to predict recurrence.Improved recurrence prediction is also linked in the passage to patient anxiety and pathological assessment.Source: B. Breast Cancer · p. 2
C. Weighted Association Rule Mining
The section contrasts conventional frequency-based rule assessment with weighted approaches and reviews WCBA as a response to the difficulty of assigning item weights.
C. Weighted Association Rule Mining
- Traditional rule mining commonly uses support and confidence while treating all items as equally important.The passage contrasts this with approaches that also consider qualitative factors such as human interpretation.Source: C. Weighted Association Rule Mining · p. 2
- WARM assigns database items weights reflecting their significance and incorporates those weights into weighted support to focus on important associations.The weights are generated using domain knowledge before being supplied to the WARM algorithms.Source: C. Weighted Association Rule Mining · p. 2
- Because domain experts may not be able to assign precise weights across many items, J. Alwidian introduced WCBA as an extension of CBA with a weighting algorithm [8].The passage notes that semi-automated or automated weighting may be more practical in such cases.Source: C. Weighted Association Rule Mining · p. 2
D. Contributions
The paper’s contributions center on coverage-based rule groups, a data-dependent default class, and removing training instances linked to major rules during pruning.
D. Contributions
- The implementation divides rules into major and minor groups according to their coverage of training instances, with rules matching no instance classified as minor.This contrasts with methods that split rules into strong and weak groups using other criteria.Source: D. Contributions · p. 2
- After partial pruning, the default class is whichever of benign or malignant has more remaining training instances, and conflicting rules are removed.Source: D. Contributions · p. 2
- Partial pruning removes training instances linked to major rules, which the paper says reduces rule redundancy and improves resource efficiency.Source: D. Contributions · p. 3
E. Paper Structure
The paper previews its implementation and experiment, then reviews prior association-rule and machine-learning approaches to breast cancer detection, diagnosis, and recurrence prediction.
E. Paper Structure
- The implementation is described in section 3, while section 4 covers the dataset and visualizes class and feature distributions.The paper also says section 2 analyzes related work and its limitations.Source: E. Paper Structure · p. 3
II. RELATED WORK
- Alwidan et al. introduced WCBA, using expert-assigned attribute weights, dual classification, and a statistical measure for rule generation.Source: §II · p. 3
- Pala et al. reported a 100% nonrecurrence rate for a rule involving absent lymph-node capsules, while Umesh et al. reported a minimum classification error of 0.0724.Source: §II · p. 3
- Other reviewed work applies FP-growth to classify breast masses, uses association rules for risk-factor assessment, and explores gene–transcription-factor rules and improved Apriori mining.Source: §II · p. 3
- Karabatak et al. reported higher accuracy from combining association rules with neural networks than using neural networks alone, and Palaniappan reported a 98.4% correct rate.Source: §II · p. 3
- Prior BC datasets have been identified as needing additional relevant attributes and better missing-value handling for recurrence prediction; Bayesian-network integration is proposed for future work.Source: §II · pp. 3–4
III. AR FOR BREAST CANCER CLASSIFICATION
The breast cancer classification implementation consists of rule generation, pruning, and prediction algorithms.
III. AR FOR BREAST CANCER CLASSIFICATION
- The association-rule classification implementation assigns rule generation, pruning, and prediction to three main algorithms.Source: §III · p. 4
A. Rule Generation
Rule Generation identifies frequent itemsets and derives class-specific association rules using weighted support, with attribute weights based on domain experts’ WCBA approach.
A. Rule Generation
- The algorithm counts itemsets by class, calculates weighted support, filters by minimum support, and generates new itemsets from frequent ones.Source: A. Rule Generation · p. 4
- Rule Generation iteratively identifies frequent itemsets and generates separate if-then rules for benign and malignant classes using a join-and-prune strategy.Source: A. Rule Generation · p. 4
- Attribute weights are predefined for selected dataset columns, based on WCBA weights obtained from domain experts at King Hussein Cancer Center.Preprocessing drops unnecessary columns, converts the dataset into lists, and applies the weights.Source: A. Rule Generation · p. 4
B. Rule Pruning
Rule Pruning filters class-specific rules by confidence and harmonic mean, resolves conflicts, and partitions rules according to their coverage of training instances.
B. Rule Pruning
- Rule Pruning first applies a minimum-confidence threshold, then calculates each rule’s harmonic mean and sorts rules by that value.Source: B. Rule Pruning · p. 4
- Conflicting benign and malignant rules are deleted before partial pruning separates major and minor rules based on their presence in training instances.Source: B. Rule Pruning · p. 4
- Partial pruning adds rules covering at least one training instance and removes linked instances iteratively; unmatched rules are minor, and remaining instances determine a default class by majority.Source: B. Rule Pruning · p. 4
C. Rule Prediction
Rule Prediction classifies test instances by comparing matched benign and malignant rules, while the experiments report correct predictions on selected samples despite dataset imbalance.
C. Rule Prediction
- For each test instance, Rule Prediction matches benign and malignant rules, compares their harmonic-mean averages, and uses a default class if needed.The algorithm uses major and minor rules produced by earlier stages.Source: C. Rule Prediction · p. 4
A. Dataset
- The dataset has 699 instances and 10 attributes, with benign samples far outnumbering malignant samples and potential effects on generalization and rule generation.The authors say they could not address the imbalance through data augmentation, which medical experts considered unacceptable or very difficult.Source: A. Dataset · pp. 4–5
IV. EXPERIMENTS
- After generating and pruning rules, the authors tested selected dataset samples and report that the prediction algorithm correctly classified them as benign or malignant.The supplied Table I caption identifies testing results and evaluation metrics, but does not provide their values.Source: A. Dataset · p. 5, Table I · p. 5
B. Experimental Setup
The experiments evaluate the classifier using accuracy and F1 across input-size categories, support values, and a fixed confidence setting on a specified machine.
B. Experimental Setup
- Training and testing ran on a Linux Ubuntu virtual machine with 4 GB of RAM and an AMD Ryzen quadcore processor.Source: B. Experimental Setup · p. 5
- Testing used accuracy and F1, with attributes grouped into three input-size categories and support values of 10%, 20%, and 30% at 50% confidence.The categories contain one to three, four to six, and seven to nine input attributes, respectively.Source: B. Experimental Setup · p. 5
- The test data comprised 140 samples; missing values were replaced with zeroes, and random zeroes were added to samples to meet category-specific input counts.The nine attributes were discretized from 1 to 10.Source: B. Experimental Setup · p. 5
C. Experimental Results
The reported support setting associated with the strongest accuracy and F1 differs by input category, and the authors describe the implementation as correctly predicting tested labels.
C. Experimental Results
- Category 1 and Category 3 had their highest accuracy and F1 at 20% support, while Category 2 performed best around 10%.For Category 1, raising support from 10% to 20% improved both metrics, while 30% lowered them below their original values; Category 2 metrics declined as support increased.Source: C. Experimental Results · pp. 5–6
- Across categories, increasing support generally lowered or stabilized performance, but the pattern was not consistent across all categories and metrics.Source: C. Experimental Results · p. 6
- The authors report that the prediction algorithm correctly classified the tested samples and describe the implementation as an Association Rules Classifier for breast cancer.They characterize it as building on weighted classification and introducing custom methods, including default-class selection and major/minor rule division.Source: CONCLUSION · p. 6