B2Classifier –一個適用於生物學資料之二元分類的整合型機器學習網站

Abstract

B2Classifier是一個用來協助使用者建立二元分類器的建模網站。它具備簡易操作、免完裝且可合併多類不同訓練模型的特性，能讓不具備撰寫程式能力的使用者亦得以利用此輔助建模工具建立二元分類器。它提供四種訓練模型，分別是artificial neural network (ANN), support vector machine (SVM), random forest (RF)和hierarchical feature integration procedure (HI)。其中 ANN 訓練模型是利用Fast Artificial Neural Network Library實作，使用者可以利用簡單的網頁畫面，勾選不同參數來建構出ANN模型。SVM是以LIBSVM完成，RF則是利用C4.5演算法建構。最後一項HI訓練模型，是透過建立IFtree來進行預測。HI是一種讓使用者以自身知識來進行特徵選擇的方法，不同以往的特徵選擇法是使用近似最佳法。HI讓使用者以其專業知識對特徵做分群，接著自動化地透過粒子群演算法 (PSO, particle swarm optimization) 來決定每類特徵群的權重。IFtree依據每個節點上的權重計算出IF (integrated feature) 值，最後利用根結點的IF值進行預測。

B2Classifier is a web system for creating binary classifiers. It has a user-friendly interface. Users do not need to install any particular software before using B2Classifier. In addition, B2Classifier provides several different machine learning methods and the classification powers of these methods can be easily integrated. Four machine learning methods are implemented in the B2Classifier, including the artificial neural network (ANN), support vector machine (SVM), random forest (RF) and the hierarchical feature integration procedure (HI). The ANN method is implemented by utilizing the Fast Artificial Neural Network Library. Users can use a simple graphic interface to set parameters and create ANN models. The SVM module of B2Classifier recruits the LIBSVM package. The RF method is implemented based on the C4.5 decision tree package. The HI method relies on an IFtree to perform classifications. HI is a knowledge-based machine learning method and is very different from traditional heuristic methods. It allows users to group input features according to their professional knowledge about the features, and then B2Classifier automatically uses a PSO (particle swarm optimization) algorithm to integrate the grouped features with optimized weights.