Origin–Destination Matrix Estimation and Prediction from Socioeconomic Variables Using Automatic Feature Selection Procedure-Based Machine Learning Model

《Origin–Destination Matrix Estimation and Prediction from Socioeconomic Variables Using Automatic Feature Selection Procedure-Based Machine Learning Model》

打印

作者单位: Civil Engineer, Technical Dept., Metropolitan Transport Consortium of Campo de Gibraltar, Algeciras 11207, Spain (corresponding author). ORCID: https://orcid.org/0000-0001-9306-1168. Email: [email protected];Doctor, Dept. of Industrial and Civil Engineering, Polytechnic School of Engineering, Univ. of Cádiz, Algeciras 11202, Spain. ORCID: https://orci.org/0000-0002-2170-0693. Email: [email protected];Computer Engineer, Dept. of Computer Science Engineering, Polytechnic School of Engineering, Univ. of Cádiz, Algeciras 11202, Spain. ORCID: https://orcid.org/0000-0002-5765-369X. Email: [email protected];Doctor, Dept. of Computer Science Engineering, Polytechnic School of Engineering, Univ. of Cádiz, Algeciras 11202, Spain. Email: [email protected]

摘要: The origin–destination (OD) demand matrix plays an essential role in travel modeling and transport planning. Traditional OD matrices are estimated from expensive and laborious traffic counts and surveys. Accordingly, this study proposes a new combined methodology to estimate or update OD matrices (urban mobility) directly from easy-to-obtain and free-of-charge socioeconomic variables. The Málaga region, Spain, was used as a case study. The proposed methodology involves two stages. First, an automatic feature selection procedure was developed to determine the most relevant socioeconomic variables, discarding the irrelevant ones. Several feature selection techniques were studied and combined. Second, machine learning (ML) models were used to estimate mobility between predefined zones. Artificial neural networks (ANNs) and support vector regression (SVR) were tested and compared using the most relevant variables as inputs. The experimental results show that the proposed combined model can be more accurate than traditional methods and ML models without the feature selection procedure. In particular, SVR with feature selection slightly outperformed the combined model using ANNs. The proposed methodology can be a promising and affordable alternative method for estimating OD matrices, reducing costs and lead time significantly, and assisting and improving urban transport planning.