基於熵的特徵選擇、複數模糊和混合機器學習於預測研究的新方法

A Novel Approach to Prediction Problems with Entropy-Based Feature Selection, Complex-Valued Fuzzy Method, and Hybrid Machine Learning

大數據中大量特徵帶來高運算成本和風險，減緩資料維度與參數量對模型影響是重要的課題。因此，本研究提出新的全複數模糊模型，以複數型輸入減少資料維度，進而降低計算複雜度與時間成本，並以複數型輸出達到雙目標預測。其次，基於熵來改良多目標特徵選擇演算法，提供模型更有力之特徵、以輸入空間最適區塊挑選法降低參數量，並於機器學習中提出新混合式演算法，混合高斯鯨群演算法更新模型前鑑部參數及遞迴最小平方估計法更新後鑑部參數，稱作GD-WOARLSE（Gaussian Distribution based Whale Optimization Algorithm with Recursive Least Squares Estimator）。最後以三個實驗驗證模型效能，包含MackeyGlass時間序列、同時逼近兩個函數與美國兩個常用股票指數S&P500與道瓊工業指數，在評估模型特徵選擇、雙目標預測和參數減少影響均有不錯表現。

With the rise of big data, the increase in the number of features has led to higher computation costs and larger parameter risks in models. It is a crucial issue to attenuate the effects of input dimensionality and parameter quantity under limited resources. This study proposes a new complex fuzzy model, which can accept complex input to reduce data dimensionality, thereby lowering model computation complexity and time costs, and achieving dual-target prediction with complex-valued model output. Based on the concept of entropy, this study improves the multi-target feature selection algorithm and attenuates the number of parameters through the method of grid-type selection of input space. Additionally, a Gaussian distribution based whale optimization algorithm with recursive least squares estimator (GD-WOA-RLSE) hybrid algorithm is proposed to update the parameters of the fuzzy model in machine learning. GD-WOA is employed to update the parameters of the Ifparts, while RLSE is utilized to update the parameters of the Then-parts. The performance of model was evaluated in three experiments, including the Mackey Glass time series, simultaneous approximation of two functions, and prediction of two commonly used US stock indices, S&P500 and Dow Jones Industrial average (DJI). The evaluation shows good performance in terms of feature selection, parameter reduction, and dual-target prediction ability.