Optimizing machine learning methods for trajectory data pattern extraction and traffic predictions

Authors: Xuantong Wang*,
Topics: Cyberinfrastructure, Transportation Geography, Geographic Information Science and Systems
Keywords: machine learning, GIS, AIS, GPU
Session Type: Paper
Day: 4/4/2019
Start / End Time: 5:00 PM / 6:40 PM
Room: 8228, Park Tower Suites, Marriott, Lobby Level
Presentation File: No File Uploaded


The increasing availability of GPS-enabled devices have produced a massive amount of spatial trajectory data to represent the mobility of various moving objects. Performing data mining on this massive data can help us extract useful information, patterns, and trends that were previously unknown. Scientists and practitioners have proposed numerous spatiotemporal machine learning (ML) analytical models to analyze trajectory data. Nevertheless, these ML models suffer from many deficiencies that can significantly limit their applications. First, these generic methods do not incorporate geospatial knowledge and principles when formulating the models. Second, these methods do not consider the data processing requirements and thus fail to provide efficient spatiotemporal data processing capabilities for massive data. Hence, the aim of this study is to address these challenges by 1) including a set of spatial temporal constraints to customize the configuration of the methods, 2) developing an integrated and interactive visualization platform through which users can fine tune the models to improve accuracy time series forecast and pattern extractions, and 3) providing support for spatiotemporal data analysis by utilizing the many-core computing power of GPU-based parallel data management and processing. In this study, we implement several a set of popular trajectory data mining methods in a parallel manner to compare performance between traditional and our optimized ML methods. Results show that our GPU-based ML methods can significantly improve the performance accuracy and speed up data processing. Customizable functions also effectively allow users to fine tune the ML analysis based on real-time performance feedback.

Abstract Information

This abstract is already part of a session. View the session here.

To access contact information login