An-Syu Li

This study analyzes historical data to identify three key factors that play critical roles in human mobility patterns. Based on these factors, a temporal-trajectory-based K-nearest neighbor algorithm was developed to predict human flow trajectories. The algorithm demonstrates effectiveness when applied to the HuMob Challenge 2024 dataset, showing how spatial-temporal patterns and individual mobility behaviors can be leveraged for accurate trajectory prediction in urban environments.

This research applies generative pre-trained transformer (GPT) models to predict traffic accidents in Taiwan. By leveraging the natural language processing capabilities of transformers and fine-tuning on traffic accident data, the study demonstrates how large language models can be adapted for predictive modeling in transportation safety. The approach shows promise for understanding complex patterns in accident data and providing actionable insights for traffic management and accident prevention strategies.

This paper presents a novel approach to personalized travel recommendation by mining social media itineraries and applying spatial discretization techniques. The system extracts travel patterns from user-generated content and uses spatial analysis methods to create customized travel suggestions. By combining collaborative filtering with geographic information systems, the research demonstrates how social media data can be leveraged to understand travel preferences and generate location-aware recommendations that match individual user interests.

This research introduces a deep learning-based travel recommendation system that utilizes heatmap visualization to represent spatial patterns of tourist activity. By training neural networks on geospatial data, the system generates visual representations of popular destinations and personalized recommendations based on user preferences and historical travel patterns. The heatmap approach provides an intuitive interface for exploring travel options while maintaining the sophistication of deep learning models.

This paper explores the use of synthetic telecommunications data for building community resilience applications. The research addresses privacy concerns while maintaining data utility by generating synthetic datasets that preserve the statistical properties of real telecom data. The study demonstrates how synthetic data can enable the development and testing of emergency response systems, disaster management tools, and community connectivity applications without compromising individual privacy.

MulSAFER introduces a comprehensive framework for extracting and interpreting safe mobility routes from multimodal spatial data sources. The system integrates data from traffic sensors, accident reports, infrastructure information, and environmental factors to identify the safest paths for various transportation modes. By combining machine learning with geographic analysis, MulSAFER provides interpretable recommendations that balance safety considerations with travel efficiency.

This innovative research applies conditional adversarial generative networks (CGAN) to automatically generate urban intersection road markings in different national styles from aerial imagery. The system learns the characteristic patterns of road markings across various countries and can synthesize appropriate markings for new intersections while adhering to local standards. This approach has significant implications for automated mapping, urban planning, and infrastructure development.

This work presents a novel framework combining DeBERTa (Decoding-enhanced BERT with disentangled attention) with Conditional Generative Adversarial Networks for privacy-preserving synthetic data generation. The framework focuses on generating synthetic text data that preserves the semantic and statistical properties of original datasets while protecting individual privacy. The research demonstrates robust performance on natural language inference tasks, showing that synthetic data can effectively replace real data in machine learning applications without compromising model accuracy.

This research presents GeoChronos-GPT, a novel spatiotemporal transformer framework designed for personalized point of interest (POI) recommendation. The framework leverages transformer architecture to capture complex temporal patterns and spatial dependencies in user mobility data, enabling accurate prediction of future visit locations. By combining the sequential modeling capabilities of GPT-style transformers with geospatial analysis techniques, GeoChronos-GPT provides context-aware recommendations that account for both individual user preferences and broader spatiotemporal trends in location-based services.

Integrating style transfer and recommendation system implementations to deliver personalized travel recommendations.

Implemented a Temporal-Trajectory-based K Nearest Neighbor Algorithm to predict human mobility patterns based on historical spatial data.

Using statistics and user-based collaborative filtering to implement a recommender system that effectively predicts the later few songs listened to by users.