Human-in-the-loop Machine Learning Framework for Discovering Interesting Association Patterns in Space-Time Big Data

Authors: Arif Masrur*, Pennsylvania State University
Topics: Geographic Information Science and Systems
Keywords: Big Geodata, machine leanrning, data mining, wildfires
Session Type: Paper
Day: 4/5/2019
Start / End Time: 8:00 AM / 9:40 AM
Room: Washington 6, Marriott, Exhibition Level
Presentation File: No File Uploaded


Discovery of interesting spatio-temporal patterns in large, multi-dimensional, event-based datasets offers a new methodology to aid in making important decisions during many real-world problems i.e. wildfires, spread of diseases, etc. However, pattern mining algorithms usually generate an overwhelmingly large number of patterns that cannot be sifted manually to find interesting ones according to a particular user’s domain knowledge and objectives. In addressing this important problem, data mining research proposes different user feedback-based methodological frameworks. While existing approaches have been shown to work well for interesting pattern discovery for discrete event datasets, e.g. market transactions, the one-step user feedback mechanism these employ can be ineffective when dealing with the problem of interesting pattern discovery in continuous, scale-dependent, and multivariate spatio-temporal event datasets. Based on a semi-supervised learning technique called self-training, the proposed research aims to develop a methodological framework that can support a two-step user-feedback mechanism at each iterative session: one that facilitates modeling of the user’s subjective interestingness preference of patterns and other helps to diagnose model’s performance, when necessary. This iterative two-step knowledge-based guidance is assumed to yield an improved interestingness model for detecting truly interesting patterns that are better suited for complex spatio-temporal data. Multiple user evaluations will be conducted using wildfire and related biophysical data to validate and evaluate the performance of the proposed methodology in discovering interesting spatio-temporal patterns of lesser-known dynamics of recent wildfire events in the North American regions (e.g. Alaska, Canada, California).

Abstract Information

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

To access contact information login