GPS-based travel survey is an emerging technology proved to be effective for trip chain information collection. Existing rule-based algorithms for trip information detection have limitations in accuracy, versatility and portability. Besides, the influence of several key technical factors in application, such as travel mode, traffic condition, data sampling frequency and data processing algorithms etc. have not been analyzed and evaluated. Therefore, in this paper, a hybrid model for GPS-based travel survey is proposed based on the performance evaluation and comparison using different methods and data. First, four most popular machine learning algorithms (MLAs) including neural network, support vector machine, Bayesian network and random forest, cooperated with a GIS-based map matching algorithm (GMM), are used to extract trip chain information; Second, the influence of different technical factors including trip mode (10 multi-modes), data sampling frequencies (1s to 120s), traffic conditions (non-peak and peak hour traffic) and algorithms (only MLAs and MLAs+GMM) are evaluated. Results show that all the proposed algorithms can be applied for GPS-based trip mode detection. Performances are similar and relatively low when using only MLAs. The GMM algorithm contributes a lot to improve the bus and car mode detection. Data sampling frequency and traffic condition obviously influence the model performance. A high data sampling frequency and free traffic condition helps to improve the outcomes. Trip mode detection rates reach 80% and mode transfer time detection errors are within 1 minute when data sampling frequency is smaller than 5s both under free and congestion condition. Keywords: GPS-based Travel Survey, Machine Learning, GIS, Data Sampling Frequency, Traffic Condition

CICTP