Reduction and Clustering of Eye-movement Data

[[abstract]]人類的決策行為可以透過各種不同的設備而取得觀察資料，而眼動儀是一種紀錄眼睛注視軌跡的儀器，許多研究會透過設計不同的決策策略來觀察與研究人類眼睛在面臨決策下的移動方式。透過眼動儀，我們將會得到大量的決策資訊，倘若我們要將一筆新加入的眼動資料分類，就必須比對大量過去的眼動資料來得知新加入的資料是屬於哪一種決策模式，隨著歷史資料的增加，往後所需要的比對時間也會隨著增長。因此如何將眾多的歷史資料壓縮成具有代表性的資料就變得十分重要。此外，若我們一開始無法得知一大群眼動資料中包含什麼屬性，如何將這些眼動資料有效的分群也會變得相當重要。 本研究將提出一個使用混合式的多重啟始禁忌搜尋法(Hybrid-MTS)來解決眼動資料的壓縮與分群問題。在壓縮問題中，必須確保壓縮前的辨識率與壓縮後的辨識率誤差在一定的範圍之內。而在分群問題中，希望最小化每一群內的距離總和，以符合分群的精神。而本研究將透過真實資料來驗證所提出的演算法是否適用於眼動資料的處理。[[abstract]]The observed data for capturing human decision behaviors can be obtained by various devices, among which, eye trackers are frequently used to collect fixation data. Many researches study the eye-movement data under different decision strategies. Given a new piece of eye-movement data, we need to classify it as one of the decision strategies by matching with historical data. To save the matching time, it is thus very important to reduce the historical data. Moreover, it is also important to cluster the historical data if they are not labeled. In this study, a hybrid multi-start tabu search (Hybrid-MTS) method was developed to solve the reduction and clustering problems for eye-movement data. For the data reduction problem, it is stipulated that the difference between the classification rate before and after the data reduction process should be less than a threshold. For the data clustering problem, the sum of distances between any two members within a cluster is minimized. Our proposed methods are validated by real world eye-movement data.[[note]]碩