Prof. Chung-Chian Hsu
National Yunlin University of Science and Technology, Taiwan

Speech Title: Traffic Volume Prediction via An Explainable Deep Learning Model with Variational Mode Decomposition and Multiple Temporal Features
Abstract: Accurately predicting short-term traffic flow is one of the key issues in smart city management. With the rapid development of deep learning technologies, an increasing number of researchers have attempted to apply advanced time series models, such as Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) to traffic flow prediction to capture its nonlinear dynamics and long-term dependency characteristics. However, relying solely on deep neural networks is still insufficient to fully overcome the high variability inherent in traffic data. If the noise in the data is not effectively addressed, the model may misinterpret the data structure, thereby affecting prediction accuracy and stability. As a result, data preprocessing methods have been increasingly emphasized, with Variational Mode Decomposition (VMD) being one commonly used technique. VMD can decompose the original time series signal into multiple Intrinsic Mode Functions (IMFs) with different frequency characteristics, which helps reduce noise, extract primary trends, and enhance the model's ability to understand the structure of time series data, thereby improving prediction accuracy. Moreover, although deep learning models such as LSTM and GRU possess excellent capabilities for time series data modeling, their 'black box' nature makes it difficult to explain the specific contributions of input features to the prediction results, limiting their trustworthiness in sensitive application scenarios such as public policy and resource allocation. Particularly in contexts that combine multiple feature sources (such as temporal context, historical flow, and variational mode decomposition), the relationship between model inputs and outputs becomes more complex. Without explanatory mechanisms to assist, it can hinder subsequent feature optimization and practical communication. Therefore, enhancing model interpretability, analyzing feature contributions, and clarifying the logic behind model judgments are also key bottlenecks that need to be overcome in the field of traffic flow prediction. To address these issues, this study tackles the challenges of short-term traffic flow prediction by introducing a deep learning framework that integrates multiple feature sources with variational mode decomposition and explainable artificial intelligence techniques, aiming to improve the model's accuracy and explainability. The multiple features are divided into three main modules as inputs to the prediction model: traffic and temporal information, cross-day historical data, and traffic frequency structures. The traffic and temporal module includes features of traffic flow, time period, weekday, and holiday. The cross-day historical data module consists of traffic flow data from the same time points over the past few days. The traffic frequency structure module contains frequency sequences obtained through variational mode decomposition. In terms of explainability of the predictive model, we applied SHAP (SHapley Additive exPlanations) technique. SHAP, as one of the explainable artificial intelligence techniques, has demonstrated advantages in various applications. SHAP quantifies the contribution of input features to model predictions, thereby enhancing the explainability and transparency of the model, particularly in feature impact analysis within deep learning models. This study conducted experiments using a publicly available traffic dataset from a city in central Taiwan. The results of the ablation experiments indicate that, firstly, incorporating temporal features such as "time period," "weekday," and "holiday" from the first module significantly improves prediction performance, demonstrating a high degree of complementarity. In particular, using the Mean Absolute Percentage Error (MAPE) as a metric, the baseline model that only utilized traffic flow features achieved a MAPE of 17.23. When the additional temporal information was included, the MAPE dropped to 15.33, representing an 11.03% reduction. Secondly, incorporating the cross-day historical data from the second module further enhances the model's ability to learn repetitive traffic patterns, making the predictions more stable and capable of capturing long-term dependencies. The MAPE decreased to 14.64, representing a 4.5% improvement compared to the performance achieved using only traffic and temporal features. Thirdly, when the traffic frequency structure from the third module is incorporated, the overall prediction performance is further optimized. Using the model with only the traffic and temporal modules as the baseline, integrating with the cross-day historical data module and the traffic frequency structure module reduced the MAPE to 14.47, representing a 5.64% improvement. Analysis of featuresimportance by SHAP reveals that top ten important features among the 247 input features include the time point of the prediction target (i.e., time t + 1), the time point t + 2, the indicator of holiday or not at the time point of the prediction target, the first and the second mode from the intrinsic mode functions (IMFs) of the current time point t, the first mode from the IMFs of the prior time point t – 1, the traffic volumes at the same time point t + 1 of one-day, two-day, three-day, and one-week prior to the prediction day. Note that IMFs are the time series data generated by variational mode decomposition. In general, one common characteristic of the top ten important features is that the time point of these features are either the same with or close to the time point t + 1 of the prediction target. The experimental results verify that the proposed three-module integrated framework exhibits robust error suppression capabilities, significantly enhancing the overall prediction quality and stability of the model. Furthermore, through explainable artificial intelligence analysis techniques, the quantification and visualization of feature contributions are achieved, assisting users in understanding the model's prediction logic, thereby enhancing decision-making bases and policy communication capabilities.

Biography: Chung-Chian Hsu holds a Ph.D. in Computer Science from Northwestern University, USA. Currently, he serves as a professor in the Department of Information Management at National Yunlin University of Science and Technology (NYUST) and holds the position of Director of the Information Division at the Foundation for Testing Center for Technological and Vocational Education in Taiwan. Previously, the speaker served as a Distinguished Professor at NYUST, Chair of the Department of Information Management, and Director of the International Graduate School of Artificial Intelligence at NYUST. He has received numerous accolades, including the NYUST Outstanding Research and Development Award, multiple Excellent Teaching Awards, and the National Science Council's Special Outstanding Talent Award. The speaker has collaborated on various academia-industry projects with organizations such as National Taiwan University Hospital Yunlin Branch, National Cheng Kung University Hospital Yunlin Branch, Dalin Tzu Chi Hospital, and WPG Holdings. His research interests include Artificial Intelligence, Deep Learning, Machine Learning, and Big Data Analytics. His research findings have been published in top-tier academic journals, including IEEE Transactions on Knowledge and Data Engineering, IEEE Transactions on Neural Networks, IEEE Transactions on Neural Networks and Learning Systems, and ACM Transactions on Asian Language Information Process.

Invited Speaker II

Prof. Emanuel S. Grant
University of North Dakota, USA  

Speech Title: TBA
Abstract: TBA

Biography: Emanuel S. Grant received a B.Sc. from the University of the West Indies, MCS from Florida Atlantic University, and a Ph.D. from Colorado State University, all in Computer Science. Since 2008, he is an Associate Professor in the Department of Computer Science (August 2002 – June 2018) and the School of Electrical Engineering and Computer Science (June 2018 – present) at the University of North Dakota, USA, where he started as an Assistant Professor in 2002. He currently serves as the Associate Director of the School of Electrical Engineering and Computer Science (SEECS) and SEECS Graduate Program Director. His research interests are in AI integration into software development, software development methodologies, formal specification techniques, domain-specific modeling languages, model-driven software development, software engineering education, and ethics for software engineering Emanuel Grant has conducted research in software engineering teaching with collaborators from Holy Angel University, Philippines; HELP University College, Malaysia; III-Hyderabad, India; Singapore Management University, Singapore; Montclair State University, and University of North Carolina Wilmington of the USA; and the University of Technology, Jamaica. He is affiliated with the SEMAT (Software Engineering Method and Theory) organization, as a member of the Essence - Kernel and Language for Software Engineering Methods (Essence) group. Emanuel is a member of the Association for Computing Machinery (ACM), Upsilon Pi Epsilon (UPE), and the Institute of Electrical and Electronics Engineers (IEEE).