比較分群法處理個體選擇模式中個體偏好差異之研究

Taste Variation in Disaggregate Choice Model：A Comparative Study of User Segmentation Techniques

許峰銓(Feng-Chuan Hsu)

指導教授 ： 董啟崇

淡江大學/商管學院/運輸管理學系碩士班/碩士(2018年)

https://doi.org/10.6846/TKU.2018.00246

摘要

進行旅運行為分析時，常使用市場區隔法將旅程決策者分群處理，其目的在於將性質相近的決策者歸為同一類，藉由分類的方式區分屬於不同分群的資料間之差異，使得屬於同一分群的決策者有較高的同質性，並且跟隸屬其他分群者有較多差異，此即市場區隔。建立市場區隔時需要使用到若干分群方法，包括由分析者根據先驗知識主觀假設不同群間之差異，也可藉由數學或統計方法之演算得出分群結果，本研究將選擇三種不同分群方法之分群結果予以比較分析，並根據分群結果建立之個體選擇模式進行討論。
本研究以國道五號小客車使用者為研究對象，分別藉由面訪與網路問卷蒐集，探討其運具移轉行為資料，運用三個分別具有不同特色之分群方法，包括K-means、隨機森林以及潛在類別模式進行分群，比較分群結果特徵與行為差異。研究結果顯示即使三個分群方法之最適分群數不盡相同，但影響分群效果之變數大致相同，其中18-29歲之學生與年輕族群和其他族群具有明顯區隔，且18-29歲族群在充分理解大眾運輸服務場站與路網資訊條件下，願意轉移至大眾運輸工具之比例較其他族群高。
本研究係以隨機抽取80%之觀察資料根據分群結果分別建立個別群組之羅吉特模式，其中另以旅程出發地分別建立不限出發地模式與北北基桃模式兩組模式；參數校估後之檢定結果顯示分群模式與不分群模式有顯著差異，且分群模式中變數之顯著程度與不分群模式確有不同；而比較不同兩組分群模式內部子群模式之參數亦有顯著差異，顯示市場區隔具有分群之效果。另外，相較於不限出發地模式，北北基桃模式之分群模式內部具顯著差異之變數更多。
模式參數校估結果亦顯示非年輕族群與出發地非位於雙北、基隆、桃園之受訪者對於小客車容易有起始之偏好；家中可用車輛數越多或與家人同行亦會降低移轉至大眾運輸意願；降低使用大眾運輸所需花費的時間，包含進入國道五號前、行駛於國道五號上及離開國道五號後三段時間皆可提升受訪者移轉至大眾運輸之意願；其中針對移轉意願較高之年輕族群又可透過降低國道客運票價進一步提升其意願。本研究最後以未參與模式構建之剩餘20%資料進行模式驗證以作為模式預測能力之比較，結果顯示不限出發地模式中，K-means分群方法之平均預測正確率約與不分群模式相等，但隨機森林與潛在類別分群方法則皆高於不分群模式；北北基桃模式中則是隨機森林分群方法之平均預測正確率略低於不分群模式，K-means與潛在類別模式分群方法則皆高於不分群模式。
歸納三個分群方法之特徵區分能力、運具移轉意願比例之差異、不限出發地模式與北北基桃模式之分群模式內部具顯著差異參數個數、分群模式自我驗證平均正確率及分群模式預測平均正確率，本研究認為以潛在類別模式與隨機森林分群較優於K-means分群法。

關鍵字

偏好差異 ； 分群法 ； 市場區隔法 ； 羅吉特模式 ； K-means ； 隨機森林 ； 潛在類別模式 ； 國道五號 ； 運具移轉

並列摘要

The market segmentation method is often used to distinguish and model various groups of decision makers implementing trip decisions. The purpose of user segmentation is to classify decision makers of similar nature into the same group, such that the trip makers who belong to the same group have higher homogeneity within the same group but with more differences among other groups. This process of user segmentation is also referred to as market segmentation.
To implement the user segmentation, we need to use the market segmentation techniques. The segmentation techniques can be based on prior knowledge of the analyst under the assumption of differences between different groups, or applying mathematical/statistical methods. Unlike the past study using user segmentation techniques mainly to improve model accuracy, in this study, we focused on the grouping results of different user segmentation technique on both the mode choice/switch behavior and the associated disaggregate choice models.
The passenger car users of National Freeway No. 5 were selected as our demonstrate research subjects on their mode choice behavior. Three user segmentation techniques were implemented, including K-means, Random Forest and Latent Class Model. The results showed that the outcomes of grouping via the three methods were different, but the variables with grouping effect were consistent, which showed students and young people who aged 18-29 are generally distinct from other groups. In addition, such group with higher information awareness to the information of public transportation service stations and road networks, showed higher the proportion of willingness to transfer to public transport than that of others groups.
Besides the grouping results, we investigated the mode choice modeling effect in the commonly used Logit model where 80% of the data were randomly selected from the sample data for model estimation and 20% for prediction. Two models were further specified by two bases, named as “departure-unrestricted model” and “departure–restricted model” respectively, depending on the trip departure location for each user group. Estimation results showed the significant effect of the user segmentation. In addition, departure–restricted model had more significance variables than that with departure-unrestricted model.
Further, the calibration results show that non-young groups and those whose departure locations being not from Keelung, New Taipei City, Taipei and Taoyuan were more likely to have higher preference for passenger cars, and the more the number of cars available or travelling with family members will reduced their willingness to transfer to public transit. Reducing the time spent on public transit, includes three sections, before entering National Freeway No. 5, traveling on National Freeway No. 5 and after leaving National Freeway No. 5, will increase the respondents’ willingness to transfer to public transit. Besides, the young groups with higher potential to transfer to can further enhance their willingness by reducing the price of Freeway Schedule Bus Service.
Finally, the model verification were carried out with the remaining 20% of the data which weren’t used to calibrate the models. The results showed that the average prediction rate of the K-means-grouping models was approximately equal to that of the pooled model in the departure-unrestricted model, while the Random Forests-grouping models and the Latent Class Model models were both higher than the pooled model. In the departure-restricted model, the average prediction rate of the Random Forests-grouping models was about 2% lower than pooled model, while the K-means-grouping models and the Latent Class Model models were both higher than the pooled model.
As demonstrated by the findings of this study, comparing the difference of the transfer willingness ratio, the number of parameters with significant differences within the grouping-models, and the self-verification of average right of prediction rate, and average prediction switching rate, Random Forests and Latent Class Model performed better than K-means in this study.

並列關鍵字

Taste Variation ； User Segmentation Techniques ； Market Segmentation ； Logit Model ； K-means ； Random Forests ； Latent Class Model ； National Free Way No.5 ； Modal Transfer

參考文獻

33. 高筑韻，2011，以潛在群體方法模化高速公路電子自動收費系統之選擇行為，國立交通大學碩士論文

31. 范碧儒，2012，以潛在類別模式分析影響台灣中老年人之憂鬱因子，國立成功大學經濟學系碩博士班，碩士論文

29. 邱琮驊，2016，國道五號及門旅行時間對運具選擇行為影響之研究，台灣大學土木工程學研究所，碩士論文

34. 陳志華、楊子緯、張訓楨、賴永崧，2016，特徵分析和機器學習方法應用於肝臟疾病檢測，福祉科技與服務管理學刊，第四卷，第三期，第417-430頁

26. 吳建生、廖梓淋、林鈺翔，2011，兩步驟類神經網路車輛偵測器遺漏資料之填補及其應用，運輸計劃季刊，第四十卷，第一期，第1-30頁