教育、家庭背景與所得之關聯法則探勘：取證自台灣教育長期追蹤資料庫

Abstract

俗語說三分天註定，七分靠打拼，古老的智慧勉勵人們，只要肯努力都會有成功的一日，在埋頭苦幹的同時，社會中的許多例子：如政治世家、豪門後代的存在，卻像是這句古諺最好的反例，究竟成功能不能只靠打拼？先天的背景到底對後天有多大的影響？想對成功的因素有更多的認知，就不能不了解這些問題的答案。 歷年來有許多這方面的文獻研究，在經濟學跟社會學的理論架構下，經濟不平等的世代傳承是被認為存在的，換句話說，父母親的社經地位對於子代的社會成就是有顯著的關聯性，除了直接相關以外，兩者間的過渡因素也是熱門的討論題材，其中最廣為探討的因素就是教育成就。教育自古一向是華人社會的重要價值，根據教育部的統計報告指出，2015年教育支出佔台灣政府歲出五分之一，近六年佔GDP 5.2% ~ 6% 之間，高於德國、日本等先進國家，而對於教育政策修訂也是社會上十分關心的議題，傳統上認為教育在促進階級流動上有著絕對性的影響 --- 所謂「十年寒窗無人問，一舉成名天下知」，教育與社會階級的關係也一直是熱門的研究議題。 然而受限於資料收集等等研究限制，多數的相關研究只能侷限於水平時間上，因此大多數的研究皆是以學生在校活動及其家庭背景為變因，探討對於學生學習表現的影響。而在社會階級流動或是對於個人成就之變因探討方面，多數研究從社會結構分類切入，或是由父母親的教育程度、家庭收入及人口統計因素為背景，變數不能完整反應學生本人、家庭背景及學習的狀態。本文研究資料取自臺灣教育長期追蹤資料庫（Taiwan Educational Panel Survey），該計畫在 2001 年至 2007 年，每兩年進行一次、總共四波的資料蒐集，調查面向包含了學生問卷、家長問卷及老師、學校問卷，作為變因能夠完整反映學生當時的背景，同時並串接台灣教育長期追蹤資料庫後續調查（TEPS-B），以學生初入社會之工作收入狀況作為目標，探討變因對其之影響。 本研究在資料清理後，先以隨機森林演算法的分類準確率來篩選重要變因，接著對各資料庫算出非相似度矩陣，用網格搜索算出最佳權重後，以合成的非相似度矩陣進行分群，分群採用Genetic K Means演算法，確保其為全域最佳解，接者再對每一分群各自進行關聯法則探勘，同時利用k-fold cross validation交互驗證，在修剪冗餘的法則後，找出滿足相關指標的重要法則。實驗結果可以從巨觀及微觀的角度呈現，巨觀上就所有子因素抽象化為自身價值觀、家庭背景及學校生活三方面來探討，發現在學生為高二階段，自身價值觀對未來收入的影響最高，家庭背景最低。微觀上來看，將所有子因素同時考慮並對應到六種不同的情節下，以易於理解的關聯法則形式來表現。

Education is widely regarded and serving as the primary pathway to allocate the economic remuneration. Basically, every study has implied the significant relationship between education achievement and earnings. Holding this belief, most people consider education as a mechanism that can sabotage the association of economic inequality transmitted from one generation to the next. Examining the inheritance of economic privilege can be considered as a branch of issues that aims to find the determinants of one’s life achievement. Achievement typically denotes as socioeconomic status (SES) which is a combined measure based on one’s education, income and occupation in economic and sociological perspectives. The major theoretical view in economic literatures exploit the family behaviors by claiming families would transmit cultural and genetic characteristics to their offspring that effects children’s SES. Accordingly, a large number of empirical studies have discussed about relationship between family background and one’s SES. In this study, we intend to propose a mechanism to investigate the determinants that take influence on one’s economic rewards. In order to bypass the limitation mentioned, we leverage diverse data mining techniques, instead of using statistical methods solely. By utilizing association rule mining, the relationships can be easily interpreted with if-then rules. Drawing on data from Taiwan Education Panel Survey (TEPS), a longitudinal dataset held in Taiwan, one’s background in the study stage can be represented completely in aspects of family, education and self-worth which are verified as determinants of future status. Our main contribution is that we found the determinants of one’s earning would vary depending on macro or micro perspectives which are rarely mentioned in other works. In macro view, the relationship was discussed between earning and overall background of family, school and self-worth respectively. It showed that the self-worth of a student played the most important role, followed by school practice and family background. In micro view, the sub-factors were investigated in six scenarios respectively. While some of rules are supportive to the previous work as well as some are trivial, there are still unexpected and surprising insights found by our approach. These findings on one hand reflect the status-quo of our society, and on the other hand can be references in many aspects that guide us how to react for aiming a high-earning future.