影響臺灣主力農家經營休閒農業之決定因素－機械學習模型之應用

The Determinants of Operating Leisure Farm Activities among Main Farm Households-Application of Machine Learning Model

張立凡(Li-Fan Chang)

指導教授 ： 張宏浩

國立臺灣大學/生物資源暨農學院/農業經濟學研究所/碩士(2019年)

https://doi.org/10.6342/NTU201900915

摘要

鑒於近年休閒農業蓬勃發展，過去許多傳統農家戶紛紛轉型為經營休閒農業，同時政府為了解決農民所得普遍低廉以及國外廉價農產品進口之困境，因而選擇推廣休閒農業之發展，由此可見未來休閒農業對於臺灣之重要性可見一班，若能藉由此研究分析出影響農家戶選擇經營休閒農業之原因，將對未來政府推動有著實質上之幫助。
本研究將分別利用三種計量模型對此議題進行分析，分別為傳統迴歸方法以及機械學習方法之LASSO方法與CART決策樹方法，在傳統迴歸方法中當變數過多時容易產生多元共線性之問題，這也是過去文獻中時常提及的OLS方法所面對之難題，因此本研究選擇使用LASSO方法以及CART方法解決這個問題。LASSO方法透過收縮係數降低模型之維度，大幅降低了發生多重共線性的可能性；此外這兩種機械學習方法都能夠進行自動變數篩選，同時也解決了置入過多不相關變數造成的估計偏誤。
本研究透過分析102年主力農家經營調查資料，共計9,849筆樣本。根據實證結果顯示在本研究的兩階段分析中，年齡、自家男性勞力、自家女性勞力、戶長最高學歷大學以上、常僱男性員工等皆具有顯著效果，顯示出可能影響農家戶經營休閒農業意願以及經營休閒農業土地面積之可能因素。此外透過比較三種計量方法可以得知，機械學習方法的LASSO以及CART方法之均方誤差普遍低於傳統的OLS方法，因此可以推論在估計的精確度上機械學習方法優於傳統迴歸方法。透過本研究之分析，未來政府於推動休閒農業相關政策時，應更加注意農家戶是否符合此類特徵，如：年齡較高、戶長最高學歷大專以上以及自家勞動人力以及常僱男性較多等因素。此外機械學習方法在經濟分析的應用上可能會出現挑選出之變數與普遍認知並不相符的情形，因此在使用機械學習方法時應特別注意變數篩選的結果。

關鍵字

休閒農業 ； 線性機率模型 ； LASSO ； CART

並列摘要

In view of the vigorous development of leisure agriculture in recent years, many traditional farmer households have been transformed into leisure agriculture in the past. At the same time, in order to solve the general low income of farmers and the plight of foreign cheap agricultural products, the government chooses to promote the development of leisure agriculture. The importance of Taiwan can be seen in a class. If we can analyze the reasons that affect farmers' choice to operate leisure agriculture, it will have substantial help for future government promotion.
This study will use three kinds of econometric models to analyze this topic, respectively, the traditional regression method and the LASSO method and CART decision tree method of mechanical learning method. In the traditional regression method, when there are too many variables, the problem of multi-collinearity is easy to occur. It is also a problem faced by the OLS method that is often mentioned in the past literature, so this study chose to use the LASSO method and the CART method to solve this problem. The LASSO method reduces the dimension of the model by shrinking the coefficient, greatly reducing the possibility of multi-collinearity; in addition, both mechanical learning methods can perform automatic variable screening, and also solve the estimation error caused by placing too many irrelevant variables. .
This study analyzed 9,102 main farm management survey data, a total of 9,849 samples. According to the empirical results, in the two-stage analysis of this study, age, male labor, female labor, top-ranking university graduates, and regular male employees all have significant effects, indicating that it may affect farmers' willingness to operate leisure agriculture. And the possible factors for operating the area of recreational agricultural land. In addition, by comparing the three measurement methods, it can be known that the mean square error of the LASSO and CART methods of the mechanical learning method is generally lower than the traditional OLS method. Therefore, it can be inferred that the mechanical learning method is significantly superior to the traditional regression method in estimating the accuracy. Through the analysis of this study, in the future, when the government promotes the policies related to leisure agriculture, it should pay more attention to whether the farmer households meet such characteristics, such as: higher age, the highest-level college graduates, and the self-employed manpower and the number of employed men. In addition, the mechanical learning method may have a situation in which the selected variables do not conform to the general cognition in the application of economic analysis. Therefore, special attention should be paid to the results of the variable screening when using the mechanical learning method.

並列關鍵字

Leisure agriculture ； LPM ； LASSO ； CART

參考文獻

王國恩，2017。「探討農民組織、農產品銷售通路、生產標章、契作與農業所得之關聯性—來自臺灣的實證分析」。碩士論文，國立臺灣大學生物資源暨農學院農業經濟學研究所。

江志民、鄭碩亮、謝鑫能，2006。「全國休閒農業金額推估之分析」，『數據分析』，1卷1期，59-75。

江羿臻、林正昌，2014。「應用決策樹探討中學生學習成就的相關因素」，『教育心理學報』，45卷3期，303-327。

行政院主計總處，2013。「102年主力農家經營概況調查報告」。台北：行政院主計總處。

行政院農業委員會，2004。「93年休閒農業場家面性調查」。台北：行政院農業委員會。