臺灣博碩士論文加值系統

不平衡資料的分類是許多實際應用中常見的問題，當其中一類的實 例數量遠少於另一類，這個問題可能導致訓練模型出現偏差，對於預測 少數類別的結果會相對較差。由於傳統機器學習分類器的設計初衷主要 是處理相對平衡的資料集，因此它們通常無法有效地對不平衡資料集進 行分類，這類問題在相當多的研究領域中都非常重要。過去提出各種技 術來解決這個問題，例如針對多數類進行減少的欠採樣(Under-sampling) 方法，對少數類進行過採樣(Over-sampling)的方法，在欠採樣技術較常見 的為隨機欠採樣（Random Under-Sampling）這種方法為隨機地從多數類 樣本中刪除一些樣本，使其與少數類樣本數量相等，而過採樣部分則是 以擁有較多變體的 SMOTE（Synthetic Minority Over-sampling Technique） 為代表，其基本思想是通過人工合成一些少數類樣本來增加訓練集中的 少數類樣本數量，從而平衡資料集。
近年來，為了解決傳統採樣方法在不平衡資料分類中的限制，部分 結合多種採樣技術以實現更好的類分布與信息保留平衡的方法已被提出， 本研究中提出透過粒子群最佳化演算法(Particle Swarm Optimization, PSO) 結合 SMOTE 及 ENN (Edited Nearest Neighbor)的混合採樣方式來解決不 II 平衡資料分類的中由於多數類別的樣本數量較多，分類模型會傾向於將 未知樣本分類為多數類別，而忽略了少數類別的問題。
本研究中提出的方法使用公開資料集 KEEL(Knowledge Extraction based on Evolutionary Learning)來進行測試並與其他幾種常見的採樣方法 進行比較，其中會以 F1、G-MEAN 等幾種評估分數來測試，測試為在資 料集中隨機提取 24 種類別並分開運行，結果在與傳統資料級別方法、啟發 型及混合型方法、演算法型及成本敏感型方法等進行比較後有取得最高為92%， 整體平均為 80.17%的勝率表現，但是相對於進行比較的傳統方法的，啟發 式方法進行優化後的採樣在運行時間上勢必會花費更長的時間，相對來 說的好處就是，在取得最佳成結果之前需要人為處理的部分會減輕，只 須設定好初始參數就行，不需要不斷手動調整以達最佳成果。

The classification of imbalanced data is a common challenge in many practical applications where the instance count of one class is significantly lower than the other. This issue can lead to bias in training models, resulting in poorer predictions for the minority class. Traditional machine learning classifiers are primarily designed to handle relatively balanced datasets, making them often ineffective in classifying imbalanced data. This problem is crucial across various research domains. One common under-sampling technique is Random Under-Sampling, where instances are randomly removed from the majority class to match the minority class's instance count. On the other hand, oversampling is exemplified by Synthetic Minority Over-sampling Technique (SMOTE), which involves artificially generating synthetic instances for the minority class to balance the dataset.
Recent years have seen the emergence of hybrid methods that combine multiple sampling techniques to achieve better class distribution and balanced information retention, aiming to overcome the limitations of traditional sampling methods in imbalanced data classification. This study proposes a IV hybrid sampling approach that combines the Particle Swarm Optimization (PSO) algorithm with SMOTE and Edited Nearest Neighbor (ENN). This approach addresses the issue that arises due to the predominant count of instances in the majority class, which can lead the classification model to favor classifying unknown samples as the majority class, thereby neglecting the minority class.
The efficacy of the proposed methodology is evaluated by conducting experiments on the publicly accessible KEEL (Knowledge Extraction based on Evolutionary Learning) dataset,. The comparison involves several commonly used sampling methods, and the evaluation metrics employed include F1-score and G-MEAN. The experimental procedure entails randomly selecting 24 classes from the dataset, followed by the execution of individual runs. The findings demonstrate a success rate of 92%, with an average of 80.17% overall, compared to traditional data-level, heuristic, hybrid, algorithmic, and costsensitive approaches. However, it is important to note that the optimized sampling using heuristic methods, while promising better results, might consume more runtime than traditional methods. The advantage lies in reducing the need for constant manual tuning to achieve optimal outcomes, as the optimization process mainly involves setting the initial parameters appropriately.

摘要 I
Abstract III
致謝 III
目錄 VI
表目錄 VIII
圖目錄 IX
第一章　緒論 1
1.1 前言 1
1.1.1 何謂資料不平衡 1
1.1.2 資料不平衡可能遇到的問題 2
1.2 研究動機和目的 3
1.3 論文架構 3
第二章 文獻探討與技術分析 5
2.1 資料層級方法 5
2.1.1 合成少數類過抽樣技術 5
2.2.2 編輯最近鄰 7
2.2 演算法層級 7
2.3 啟發式演算法 8
2.2.1 粒子群演算法 8
2.4 成本敏感學習 12
第三章 研究方法介紹 14
3.1 基於粒子群演算法優化的SMOTE 17
3.2 基於粒子群演算法優化的ENN 19
第四章 實驗結果與分析 23
4.1 實驗目的 23
4.2 資料集介紹 23
4.3 評估方法 25
4.4 評估指標 25
4.5 參數選擇與比較 27
4.6 比較結果 29
第五章 結論與未來展望 45
參考文獻 46

[1]H. He, "Learning from Imbalanced Data," 2009.
[2]曾正豪, "乳癌人數飆最多、年輕化！　醫籲30歲以上女性篩檢乳房," 2021.
[3]wiki, "survivorship bias," 2012. [Online]. Available: https://zh.wikipedia.org/zh-tw/survivorship%20bias.
[4]A. Koren and S. A. McCarroll, "Random replication of the inactive X chromosome," Genome research, vol. 24, no. 1, pp. 64-69, 2014.
[5]N. V. Chawla, K. W. Bowyer, L. O. Hall, and W. P. Kegelmeyer, "SMOTE: synthetic minority over-sampling technique," Journal of artificial intelligence research, vol. 16, pp. 321-357, 2002.
[6]C. Cooper, A. Frieze, and J. Vera, "Random deletion in a scale-free random graph process," Internet Mathematics, vol. 1, no. 4, pp. 463-483, 2004.
[7]F. Leisch, "A toolbox for k-centroids cluster analysis," Computational statistics & data analysis, vol. 51, no. 2, pp. 526-544, 2006.
[8]R. M. Pereira, Y. M. Costa, and C. N. Silla Jr, "MLTL: A multi-label approach for the Tomek Link undersampling algorithm," Neurocomputing, vol. 383, pp. 95-105, 2020.
[9]Z. Jefferson, "Bank Data: SMOTE," 2020. [Online]. Available: https://medium.com/analytics-vidhya/bank-data-smote-b5cb01a5e0a2.
[10] H. Han, W.-Y. Wang, and B.-H. Mao, "Borderline-SMOTE: a new over-sampling method in imbalanced data sets learning," in International conference on intelligent computing, 2005: Springer, pp. 878-887.
[11] H. He, Y. Bai, E. A. Garcia, and S. Li, "ADASYN: Adaptive synthetic sampling approach for imbalanced learning," in 2008 IEEE international joint conference on neural networks (IEEE world congress on computational intelligence), 2008: Ieee, pp. 1322-1328.
[12]D. L. Wilson, "Asymptotic properties of nearest neighbor rules using edited data," IEEE Transactions on Systems, Man, and Cybernetics, no. 3, pp. 408-421, 1972.
[13] N. V. Chawla, A. Lazarevic, L. O. Hall, and K. W. Bowyer, "SMOTEBoost: Improving prediction of the minority class in boosting," in Knowledge Discovery in Databases: PKDD 2003: 7th European Conference on Principles and Practice of Knowledge Discovery in Databases, Cavtat-Dubrovnik, Croatia, September 22-26, 2003. Proceedings 7, 2003: Springer, pp. 107-119.
[14] Z. Yongqing, Z. Min, Z. Danling, M. Gang, and M. Daichuan, "Improved SMOTEBagging and its application in imbalanced data classification," in IEEE Conference Anthology, 2013: IEEE, pp. 1-5.
[15] Z. Liu et al., "Self-paced ensemble for highly imbalanced massive data classification," in 2020 IEEE 36th international conference on data engineering (ICDE), 2020: IEEE, pp. 841-852.
[16]J. Kennedy and R. Eberhart, "Particle swarm optimization," 1995, doi: 10.1109/ICNN.1995.488968.
[17]J. C. Bansal, P. K. Singh, M. Saraswat, A. Verma, S. S. Jadon, and A. Abraham, "Inertia Weight strategies in Particle Swarm Optimization," 2011.
[18]Z.-H. Zhou and X.-Y. Liu, "ON MULTI-CLASS COST-SENSITIVE LEARNING," 2010.
[19]N. V. Chawla, A. Lazarevic, L. O. Hall, and K. W. Bowyer, "SMOTEBoost: Improving Prediction of the Minority Class in Boosting," 2003.
[20]A. Bria;, C. Marrocco;, M. Molinara;, and F. Tortorella, "A ranking-based cascade approach for unbalanced data," 2012.