現今全球氣候變化極端，為維護生態及環境可永續發展，各國開始提倡節能減碳，無不希望以綠運輸的載具，作為城市交通運輸中的一環，藉此降低汽機車的使用，故公共自行車的概念應運而生。臺灣在城市在捷運和公車路網逐漸成型後，市民從捷運站及公車站到達目的地仍有一段距離，此距離走路太遠，坐車又太近。為了滿足市民最後一哩的需求，台北市政府於2009年3月推動公共自行車租賃系統YouBike，期以24小時便捷綠色運具服務取代私人運具，改善停車空間不足及尖峰時間交通壅塞情形，達到延伸大眾運輸服務範圍，突破公共運輸服務時間限制，但若站點無車可騎，將影響民眾使用意願，導致使用率降低。 本研究著重在YouBike缺車站點分析，依現有可蒐集到的開放資料（Open Data），如YouBike可借車數（每5分鐘一筆資料）、氣象局觀測資料、人事行政局國定假日行事曆，利用資料探勘監督式學習技術所分析出來的資訊，搭配BI視覺化工具，提供決策調度或執行人員運補時有所依據。 實驗方法採用Apache Spark搭配Hadoop，將資料放入HDFS後執行資料前處理，並依Spark MLlib提供的機器學習演算法，進行不同分類技術的實驗。分類技術採用Naïve Bayes、支援向量機（SVM）、隨機森林（Random Forest）等演算法，分析其結果，以試圖獲得最佳YouBike補車調度預測模型，並以視覺化方式呈現模型預測之結果以及台北市各YouBike站點當下及過去之情況。 經過實驗結果得知，針對2015/06~2016/01年的訓練集資料，隨機森林演算法的表現最佳，在其資料集來源之屬性足夠之情況下，其AUC值普遍值達到0.87左右，具有較佳參考意義。因此，本研究建議未來在進行YouBike 缺車調度預測時，可以優先採用隨機森林演算法，以此優化YouBike 調度服務作業。 ;Now weather changes dramatically all around the world. For natural ecology retention and our environment sustainable development , all countries suppose to use energy saving and carbon reduction vehicle as green transportation to become a part of city transportation. According to this concept , Public Sharing Bike was invented. After bus and MRT transportation system become mature and complete in Taiwan , there is still a certain distance between public transportation station and destination. For this last mile demand , Taipei City Government had started to implement Public Share Bike System called “YouBike” in March 2009 , hoped it can replace private vehicles , improve insufficient parking space situation and traffic congestion circumstances. On the other hand , the government also expect YouBike can extend service scope and operation time of public transportation. But the most important thing that will impact citizens intention to use public bikes is bikes shortage situations in each bike station. The dissertation focuses on shortage situation of YouBike. Using supervised machine learning algorithm with related open data , eg:YouBike availability information, weather statics ,holiday and vacation information from government schedule builds a model to predict public bike insufficient situation. The dissertation also uses a BI tool called Tableau to visualize analysis result that staffs can make bike adjustment decision according to comprehensive dashboard information. The experiment puts the source data into HDFS and uses Apache Spark to do pre-process. It also compares analysis results by three classification algorithms including Naïve Bayes , SVM and Random Forest provided by Apache Spark MLlib to try to get the best adjustment predictive model of YouBike. For being easy to acquire information , the study use Tableau to create a dashboard for presenting predictive result , past and current YouBike available situation. According to the experiment result with data of 2015/11~2016/01 YouBike availability , Random Forest is the best algorithm with average 0.87 AUC when training data set has sufficient data attributes. Therefore , the dissertation suggests that operators can use Random Forest to predict bike shortage situation for improving YouBike dispatch operation.
