以旅運軌跡相似性為基礎之信令資料插補技術：運具判別之應用

A Cellular Data Interpolation Technique based on the Similarity of Travel Trajectories: Application of Mode Classification

行動信令資料常發生漂移與缺漏，導致分析與應用之困擾。本研究基於旅運軌跡相似性，利用最長共同子序列概念，提出一套整合線性插補及分群插補，依資料缺漏長度自動調整2種插補技術權重的方法，再將插補後信令資料應用於運具判別。本研究利用4項變數進行運具預測，包含旅次速度、旅次長度、公車軌跡相似度及鐵道車站相近度，區分為5種運具，採用3個監督式機器學習分類演算法：決策樹、隨機森林與倒傳遞類神經網路。結果顯示，倒傳遞類神經網路模式及插補後資料在運具判別上優於未插補及線性插補資料，也優於其他分類演算法。

Cellular data have oscillation and data missing problems hindering their applicability. In light of the similarity on travel trajectories of most travelers, this study proposes an interpolation technique based on the similarity of travel trajectories and the longest common subsequence method, integrating linear interpolation and clustering interpolation with self-adapted weights upon the length of missing data. In applying the interpolated data to model classification, four inputs, including trip speed, trip length, trajectory similarity to the locations of railway stations, and similarity to bus trajectories, are used to classify five modes based on three supervised machine learning classification algorithms: decision tree (DT), random forest (RF), and back-propagation network (BPN). The results show that the BPN model based on the data interpolated by the proposed integrated interpolation method outperforms other mode classification models based on original or linear interpolated cellular data, suggesting the applicability of the proposed model.