應用標籤鑲嵌樹架構於解決多元分類問題

Label Embedding Tree for Multi-class Classification

林威均(Lin Wei-Chun)

指導教授 ： 周珮婷 ； 黃佳慧

國立政治大學/商學院/統計學系/碩士(2021年)

https://doi.org/10.6814/NCCU202100115

摘要

在監督式的機器學習中，多類別的分類是指具有兩個以上類別的分類任務，並把每個樣本標記為其中一個類別，由於目前較常使用的多分類方法通常都對資料母體分配有所假設，或是調參較為複雜耗時，因此想要提出一個不需要母體假設，而且調參相對容易的多分類方法。本次研究所提出的方法，透過定義並計算多類別資料中，類別標籤之間的距離矩陣，以此對類別標籤進行階層式的分群，達到拆解多元分類問題的目的，然後利用這個階層樹的架構，對未分類的樣本進行多個無須資料母體假設，基於偽概似的二元分類，最終得到分類結果。本研究將所提出的分類方法應用於不同的數據集中，並與其他常見的多元分類方法進行比較，發現在不同指標下有較高的精確度，另外，本研究更進一步利用基於相互熵篩選的變數子集合提出一個多階段分類方法，發現分類準確度在連續型的數據中有所提升。

關鍵字

機器學習 ； 多元分類 ； 多元轉二元分類

並列摘要

In supervised machine learning, multi-class classification refers to a classification task with more than two categories, and each sample is marked as one of the categories. Since the commonly used multi-classification methods usually have assumptions about the distribution of data populations, or the adjustment of hyperparameters is complicated and time-consuming, we want to propose a method that does not require a population assumption and is relatively easy to adjust hyperparameters. This proposed method dismantling multiple classification problems into binary classification problems by defining and calculating the distance matrix between the category labels in the multi-class data, making a hierarchical tree between different label to disassemble the multiple classification problem, and then based on the structure of this hierarchical tree, perform multiple pseudo-likelihood binary classification on unclassified samples, and get the classification results. In this research, the target method is applied into different data sets, and compared with other common multivariate classification methods, the accuracy and macro F1 score of our target method is quite good. In addition, we propose a multi-step method to improve the classification result with the variable chosen by mutual entropy, and the result of test dataset is indeed improved.

並列關鍵字

Machine learning ； Multi-class classification ； Multi-class to binary classification

參考文獻

Allwein, E. L., Schapire, R. E., & Singer, Y. (2000). Reducing multi-class to binary: A unifying approach for margin classifiers. Journal of machine learning research, 1(Dec), 113-141.

Anthony, G., Gregg, H., & Tshilidzi, M. (2007). Image classification using SVMs: one-against-one vs one-against-all. arXiv preprint arXiv:0711.2914.

Baloochian, H., & Ghaffary, H. R. (2019). Multi-class Classification Based on Multi- criteria Decision-making. Journal of Classification, 36(1), 140-151.

Bouazizi, M., & Ohtsuki, T. (2016, May). Sentiment analysis: From binary to multi- class classification: A pattern-based approach for multi-class sentiment analysis in Twitter. In 2016 IEEE International Conference on Communications (ICC) (pp. 1-6). IEEE.

Breiman, L., Friedman, J., Stone, C. J., & Olshen, R. A. (1984). Classification and regression trees. CRC press.