Cancer Classification with a Cost-Sensitive Naive Bayes Stacking Ensemble

Yueling Xiong, Mingquan Ye, Changrong WuComput. Math. Methods Medicine 2021 · DOI

TL;DR

This paper introduces a novel "CSNB stacking" ensemble learning method for cancer classification using gene expression data. By combining a feature selection technique with a cost-sensitive Naive Bayes meta-learner, the approach effectively handles imbalanced datasets and achieves superior classification accuracy compared to existing methods.

What Problem Does This Paper Address?

Cancer classification from gene expression data is a critical task in personalized medicine, but it's fraught with challenges. Gene expression datasets are notoriously high-dimensional, often with many irrelevant or redundant features. Furthermore, these datasets are frequently imbalanced, meaning some cancer types are far more represented than others, which can bias traditional machine learning models. Existing methods, while improving, still struggle to achieve consistently high accuracy and robustness across diverse cancer types. This paper aims to address these issues by developing a more effective and reliable classification framework.

Methods

The core of this work is a two-layer stacking ensemble learning strategy.

  • Feature Selection: Before classification, the Fast Correlation-Based Feature Selection (FCBF) method is employed. FCBF identifies and removes irrelevant and redundant genes by calculating correlation and information gain metrics, significantly reducing the dimensionality of the data.
  • Primary Learners: Four well-established classifiers are used as the "base" or "primary" learners in the first layer of the ensemble:
  • LIBSVM (Support Vector Machine)
    • KNN (K-Nearest Neighbors)
    • C4.5 (Decision Tree)
    • RF (Random Forest)
  • Meta-Learner: The key innovation lies in the meta-learner used in the second layer. Recognizing the problem of imbalanced cancer data, the authors introduce a Cost-Sensitive Naive Bayes (CSNB) classifier. CSNB is designed to be sensitive to misclassification costs, giving more weight to correctly classifying rare classes. This CSNB acts as the "meta-learner," taking the predictions of the primary learners and making the final classification decision. This combination is termed "CSNB stacking."
  • Validation: A 5x10-fold nested cross-validation strategy is used to ensure robust evaluation and prevent overfitting.

Datasets

The study evaluates the proposed CSNB stacking method on nine publicly available cancer datasets from the Kent Ridge Biomedical Dataset database. These datasets represent a variety of cancer types and data characteristics:

  • NervSys (Central Nervous System Embryonal Tumors)
  • Leukemia
  • DLBCL1, DLBCL2, DLBCL3 (Diffuse Large B-cell Lymphoma)
  • Prostate Cancer
  • Ovarian Cancer
  • Lung1, Lung2 (Lung Cancer)

These datasets vary in sample size, number of genes, and whether they represent outcome prediction, protein data, or binary classification.

Key Insights and Ideas

  • Cost-Sensitivity is Crucial for Imbalanced Data: The paper strongly suggests that simply using standard classifiers on imbalanced gene expression data leads to suboptimal results. Incorporating cost-sensitivity, as done with CSNB, is vital for improving the classification of minority classes.
  • Stacking Enhances Robustness: The ensemble nature of stacking, by combining diverse base learners, inherently provides a more robust prediction than any single classifier.
  • Feature Selection is a Necessary Preprocessing Step: FCBF significantly simplifies the classification task by reducing the number of genes, which is crucial for high-dimensional genomic data. This not only improves efficiency but also potentially enhances accuracy by removing noise.
  • CSNB as a Meta-Learner is Effective: The choice of CSNB as the meta-learner proved particularly effective in handling the inherent class imbalance present in most of these cancer datasets.

Key Results

The CSNB stacking method consistently outperformed all other tested methods across the nine cancer datasets.

  • Classification Accuracy: The proposed CSNB stacking achieved the highest classification accuracy on all nine datasets. It reached 100% accuracy on Leukemia, DLBCL1, Ovarian Cancer, and Lung Cancer2. For other datasets like DLBCL3 and Prostate Cancer, it achieved significantly higher accuracies than other methods.
  • Recall Rate: The CSNB stacking method also demonstrated the highest recall rate across all datasets, indicating its effectiveness in identifying positive cases.
  • Specificity: The method generally achieved high specificity, crucial for correctly identifying negative cases.
  • F-score: The F-score, a harmonic mean of precision and recall, was also highest for CSNB stacking, further validating its superior performance.
  • Comparison: The paper compares CSNB stacking against various single classifiers (LIBSVM, KNN, RF, C4.5, NB, CSNB) and other ensemble methods (Bagging, AdaBoost, CSKNN stacking, CSC4.5 stacking, CSLIBSVM stacking). In all metrics, CSNB stacking emerged as the top performer.

Limitations and Future Work

  • Limitations: While the paper demonstrates strong performance, it focuses on binary classification problems. The authors acknowledge that the complexity of cancer can involve multiple classes.
  • Future Work: The authors propose expanding their work to multiclass classification problems. They also suggest exploring the integration of multiple data types (beyond gene expression) to achieve more detailed and comprehensive cancer analysis.

Summary updated

BibTeX
@article{xiong2021cancer,
  author = {Yueling Xiong and Mingquan Ye and Changrong Wu},
  title = {{Cancer Classification with a Cost-Sensitive Naive Bayes Stacking Ensemble}},
  journal = {Comput. Math. Methods Medicine},
  year = {2021},
  doi = {10.1155/2021/5556992},
}
Figure