顯著差異感興趣區域探勘之方法應用於眼動資料分析

Abstract

行為科學研究包含閱讀歷程、情緒、意向、安全駕駛……等課題，眼動分析均具有高度應用價值，特別在於如何提出正確假設並有效尋找顯著差異區域實屬不易。 本顯著差異感興趣區域探勘分析之方法應用於自動化眼動資料分析主要分為三個模組，「數據有效性判定模組」可以自動找出眼動資料是否是因為頭部偏移、眼動儀校正過程，所導致之無效數據樣本判定；「數據修正模組」是基於多數眼動實驗的刺激物都會以較大的方式呈現在螢幕正中間，並假設受測者不會凝視空白區域為前提所提出之方法，該模組可自動針對較大幅度的數據偏移進行修復；「顯著差異感興趣區域分析模組」是基於各個擊破法進行感興趣區域切割與定義，自動尋找刺激物中呈現顯著差異的區域。 實驗結果顯示: 「數據有效性判定模組」中提出三項機制進行眼動資料有效性檢測，以一內含50筆的眼動資料有效性判定為例，依照不同的數據有效性判定機制花費1ms~78.8ms；眼動資料修正前的無效眼動資料分布百分率為100%，經過「數據修正模組」處理後的無效眼動資料分布百分率為0%，平均處理一張刺激物的眼動資料偏移修正約1秒鐘；「顯著差異感興趣區域分析模組」可以執行兩種統計分析(獨立樣本t檢定、成對樣本t檢定)，於兩群受測者的眼動資料找尋眼動變項(ex: 首次凝視時間、總凝視時間等)中的顯著差異感興趣區域。例如: 圖片型刺激物的實驗，32位受測者、每位看4張刺激物、4種尺寸(4格、16格、64格及256格)切割的顯著差異探勘花費4.92分鐘；影片型刺激物的實驗，5位受測者、每位看20秒的刺激物、3種尺寸(4格、16格及64格)切割的顯著差異探勘花費29.38分鐘。 本顯著差異感興趣區域探勘分析之方法應用於自動化眼動資料分析解決原先必須由研究人員從事的繁瑣、耗時之分析流程，並提供相關領域研究人員日後在進行研究時，感興趣區域定義之參考。此外顯著差異感興趣區域分析模組之感興趣區域切割與定義方式並不涉及任何主觀意識，更能符合科學研究的客觀要素。

In the eye movement experiment, it is a time-consuming and exhausted task to find the significant differences between eye movement variables and independent variables in the experiment design. First, researchers needed to determine which data sets are valid or invalid. Second, definition of Regions of Interest (ROIs) or Areas of Interest (AOIs) is done according to their research hypotheses and is not automated. Then, there is a need of importing of eye movement data into the system in order to get values for variables such as the Total Contact Time, Number of Fixation, and so forth. After which, the statistical analysis is done via statistical software, such as SPSS and Microsoft Excel. The process is very tedious and time-consuming, specially, when significant differences are not found, researches need to redefine the ROIs, start over, and do the above procedure repeatedly. In order to address the mentioned difficulties, this study discusses on the method for mining significant differences in regions of interest applied to the analysis of eye movement data. There are three modules, namely Inspection of the Validation of Eye Movement Data, Correction of Shifted Eye Movement Data, and Analysis of Significant Differences in ROIs. Inspection of the Validation of Eye Movement Data module assists researchers to distinguish which eye movement data are valid or not. Results show that it took 1 millisecond~78.8 milliseconds to check the validation of one eye movement data including 50 raw data. The Correction of Shifted Eye Movement Data module was for the stimulus to be as big as possible and centered on the screen on the assumption that the participants would not gaze at this region in order to correct the shifted eye movement data. Correction of one eye movement data took an approximate time of 1 millisecond. Divide and conquer approach was applied in the Analysis of Significant Differences in ROIs module design. Its purpose is to find all potential significant differences among ROIs. It supports two kind of statistical analysis (Two-Samples t-test and Paired t-test). It tried to find all significant differences in ROIs. There were 64 participants and were divided into two groups. Each participant gazed at 4 stimuli. In this case, it took 4.92 minutes to run this significant difference mining. In another case, there were 10 participants divided into two groups. Each participant gazed at video for 20 seconds. It took around 29.38 minutes to run the significant difference mining. Method for mining significant differences in regions of interest applied to eye movement data analysis tried to provide tools that are more convenient and replace tedious procedures. Finally yet importantly, it also provided recommendation on better definition of ROIs to the researchers.