| 摘要: | 本研究以探討在運輸規劃領域當中利用行動電話資料進行運具判斷為例。運輸規劃中最重要的幾個步驟分別為旅次產生、旅次分派、運具選擇以及交通量指派。在過去都是利用例如家戶或是路邊訪問的問卷調查方式獲得運輸規劃所需的資料，但透過問卷調查的方式通常面臨了(1)耗費大量人力、(2)高拒訪率以及(3)因受訪者記憶不完整而造成錯誤填答。近年來雖然已經嘗試利用GPS資料取代過去的調查方法，但GPS資料除了不易獲得外還容易受到建築物的因素引起遮蔽效應造成定位不準確，因此不適合將GPS資料應用於大型路網中。而行動電話資料以在運輸規劃當中成為另一個備受矚目的資料蒐集方式，它無須再額外新增設備即可自動且有效率的紀錄使用者的時空資料。因此，獲得行動電話資料的成本很低，甚至可以忽略不計。在本研究當中，我們採用了兩種監督式機器學習的方法–支持向量機(SVM)與深度神經網路(DNN)–以探討在何種特徵(旅行時間、蹤跡的發生時間、兩蹤跡間的速度、旅程中最大的速度以及平均速度)、時間區段(尖峰時段、離峰時段以及全天)、運具路線組合(公共汽車路線、汽車行走於公共汽車路線與汽車行走於非公共汽車路線)以及訓練的方法，是如何影響運具判斷的準確性。
以混淆矩陣為評估指標比較各結果的準確率，結果顯示，在(1)採用五種特徵、(2)全天、(3)公車與所有汽車路線組合與(4)支持向量機進行訓練，其結果準確率高達96.58%。不幸的是，使用五種特徵進行訓練的方式需對指定的起迄對進行實際的資料蒐集，但此種方法若是應用於大型路網上是非常昂貴的。但我們可以選擇另一種可以接受的方式，採用四種特徵(蹤跡的發生時間、兩蹤跡間的速度、旅程中最大的速度以及平均速度)進行訓練(在本研究中，準確性從96.58%降為74.21%)。透過將路網中所有的起迄對以組內差異最小化組間差異最大化的方式進行分群，在各分群中挑選一起迄對進行實際的資料蒐集並建立運具判斷模型，可將此模型應用在同一分群中的其他起迄對，再利用公車電子票證進行驗證，如此一來可以減少資料蒐集的工作量。隨著電信通訊技術的發展與進步，可以預期在不久的將來利用行動電話資料進行運具判斷的準確率會更高。
還值得一提的是，在本研究中對於行動電話飄移現象的消除提出了一種改良的方法，此方法消除了在過去文獻提到的方法中所造成缺失。
;This study takes mode inference as an example to explore the usefulness of mobile phone data in the area of transportation planning. Traffic data – consisting of activity location, origin-destination pair, mode choice and traffic assignment – are essential in transportation planning. Collecting such data via a questionnaire survey, like the home or roadside interview, have long been adopted, but are usually (1) labor intensive, (2) faced with high refusal rates of respondents, and (3) relatively inaccurate due to fade-away memory. Attempts have been made to use GPS data, but GPS data are not readily available and their levels of accuracy are apt to be affected by the shielding effect due to high-rise buildings and obstacles and, hence, are not suitable to be applied in a large transportation network. Mobile phone data, emerging as a vivid data collection method for transportation planning, can automatically and effectively record transportation planning data in time-space dimension without having to add new devices. Thus, the extra cost to retrieve this phone data is small or even negligible. For this study, we adopt two supervised machine leaning methods – support vector machine (SVM) and deep neural network (DNN) – to investigate how modal features (travel time, starting time of trace, traversal speed between traces, maximum speed, and average speed), time of day (peak hours, off-peak hours, whole day), route combinations (bus route, vehicle traversing a bus route, vehicle traversing a non-bus route), and training methods (SVM and DNN) affect accuracy in inferring transportation modes (either bus or vehicle).
The results show four factors – (1) five modal features, (2) whole day data, (3) all bus and vehicle routes combined, and (4) SVM –result in better performance than other combinations in terms of an accuracy index (96.58%) or confusion matrix. Unfortunately, modal travel time between an origin and a destination in the scenario with five modal features can only be obtained by a field survey, which is costly. A second choice (consisting of four modal features – starting time of trace, traversal speed between traces, maximum speed, and average speed) can be used at an acceptable price (accuracy decreased from 96.58% to 74.21% in our experiments). The effort involved in using this four modal feature scenario in large scale networks can be reduced further by classifying used routes between O-D pairs into groups with between-group similarity minimized and within- group similarity maximized. For each group, only one route is taken for training using field survey data and for validation using smart card data; the obtained result is applied equally to other members in the same group. With expected advances in mobile phone infrastructure and technology, higher accuracy in inferring transportation modes using mobile phone data can be anticipated in the near future.
Also worthy of mention is that a novel method for elimination of the oscillation phenomenon has been proposed in this research to correct possible mistakes made by the available methods that have appeared in the literature. |
