基於行動信令資料之使用運具判斷模式

Travel Mode Classification Algorithms based on Cellular Signaling Data

本文以旅運行為調查實際蒐集資料，建構基於行動信令資料之使用運具判斷模式，利用四項變數進行運具預測，包含旅次平均速度、旅次長度、公車軌跡相似度以及鐵道場站相近度，將子旅次區分為五項運具，分別為公車、鐵道運輸、汽車、機車與步行／自行車。本研究採用基因模糊邏輯控制與三種監督式機器學習方法（K近鄰、決策樹及隨機森林），並透過5-fold交叉驗證進行模式驗證。結果顯示，因汽機車兩種運具在預測上較難劃分，決策樹與隨機森林之準確率僅可達到60％，而鐵道運輸具有最佳之預測結果，在隨機森林模式之準確率可達到86％，決策樹則可達到83％，應可進一步蒐集具更高識別度的資訊以提升汽車與機車兩項運具之準確率。

This study aims to develop mode classification algorithms based on trips and subtrips estimated by cellular signaling data. To collect the personal privacy cellular data, this study invites voluntary users for a travel diary survey. Four inputs are used to classify mode used, including travel speed, length, bus trajectory similarity, and railway stations closeness and five modes are classified, including bus, rail, car, motorcycle and walk (bike). Four algorithms, including genetic fuzzy logic controller (GFLC), K nearest neighbor (KNN), decision tree (DT) and random forest (RF) are used to classify the mode based on 5-fold cross-validation. The validation results show that the accuracies of DT and RF can reach only 60%, due to the difficulty in classifying modes of cars and motorcycles. However, the accuracy of predicting rail mode can reach as high as 86% by RF, followed by 83% by DT. Inputs with higher discrimination should be further collected to increase the correct ratio of car and motorcycle classification.