交通場站內廣告對旅客影響之研究-以台北捷運廣告為例

在整個台北捷運場站環境中，廣告數量不斷持續增加的情形下，恐會對行經於場站中的旅客產生所謂的廣告干擾(Advertising Clutter)影響。然而目前台北捷運公司在廣告的規範上，並沒有對於廣告數量的限制設定。基於期望能在未來提供管理單位，對於交通環境中的廣告數量設定建議，本研究將初步探討個別單幅廣告對於旅客的影響。本研究工作內容包含三部分：(1)本研究將運用資訊理論(Information Theory)將廣告量化，將廣告的資訊量以(Shannon Entropy)作為控制量。(2)透過控制實驗的方式，控制單一幅廣告的資訊量層級大小，觀測受測者對於廣告的反應。(3) 建立支援向量迴歸(Support Vector Regression, SVR)模式，探討個別單幅廣告對旅客之影響。本研究建立模式分為兩個部分，核心模式為僅考慮廣告「資訊量」與「廣告干擾」的關係，另一模式為基於核心模式上加入個人對於廣告的「偏好」。另外，亦使用順序羅吉斯迴歸(Ordered Logistic Regression Model)，分別建立「偏好-主觀情緒」與「偏好-客觀情緒」模式；前者探討廣告偏好與Russell環境體驗情緒之間的關係，後者為探討廣告偏好與表情分析情緒之間的關係。歸納本研究的重要發現：(1)本研究運用支援向量迴歸建立廣告「資訊量」與「廣告干擾」的關係以及加入「偏好」變數的兩種SVR模式，雖然兩種模式預測效能與我們的預期有些許落差，但已達到可接受之預測能力(MAPE<50%)。(2)運用情感運算(Affective Computing)的表情分析，所建立的「客觀情緒」與「偏好」順序羅吉斯迴歸模式，可用於預測人對於廣告的偏好感受；並且納入偏好感受可提升支援向量迴歸對於廣告干擾值的預測能力。而在構建「主觀情緒」與「偏好」的部分，則其結果不符合等成比例發生比假設條件(proportional odds assumption)。(3)本研究將SVR預測結果進行落點分佈檢視，並依循落點分佈層級進行分群結果能有效提升SVR之預測能力。惟本研究嘗試以現有的「性別」與「捷運搭乘頻率(每週搭乘天數)」兩種外顯變數，分別以「決策樹」與「潛在類別」兩種分群法的方式進行落點集群之分群效果不佳，且依據此分群結果所建立的SVR模式之預測能力亦無顯著的提升。

關鍵字

廣告干擾；資訊量；控制實驗；支援向量迴歸

並列摘要

The number of advertisement has increased enormously in the stations of Taipei Metro System, which may cause so-called “Advertising Clutter” to Metro commuters. However, this issue has long been neglected and there is no such regulation regarding total quantity control by the management of Taipei Rapid Transit Corporation (TRTC). Due to this particular concern, this thesis proposes an initiative study on how the presence of an individual commercial advertisement affects the metro commuters. Hopefully, it may serve the foundation to total advertisement quantity control in each Metro station, where the effects of all advertisements will be accumulated. There are three major parts in this thesis: (1) To quantify each advertisement by the concept of Shannon Entropy, (2) Using controlled experiment technique to observe the response of a selected group of commuters to each single advertisement with various designated Entropy level, (3) Two variant support vector regression (SVR) models being constructed to relate the advertisement entropy to the advertising clutter with or without including individual commuter’s preference toward the advertisement. In addition, Ordered Logistic Regression Models were investigated to depict the relation between commuter’s preference to subjective affective components defined by Russell’s eight emotional experience variables and objective affective components defined by eight emotional attributes of Microsoft Emotion API using Azure Media Face Detector. The major findings of this thesis are: (1) The SVR models being successfully constructed with acceptable MAPE (less than 50%), although not ideal; (2) Preference-objective affection model being a better model than preference-subjective affection model due to the violation of proportional odds assumption by the later; and (3) Market segmentation by placement showing dramatic improvement over generic SVR model, however, efforts to construct placement clusters by Latent Preference model or decision tree technique being not successful using current data available.