Research Area:  Machine Learning
Finding Earth science data has been a challenging problem given both the quantity of data availableand the heterogeneity of the data across a wide variety of domains. Current search engines in most geospatial data portals tend to induce end users to focus on one single data attribute.This approach largely fails to take account of users multipleand dynamic preferences for geospatial data,and hence may likely result in a less than optimal user experience in discovering the most applicable dataset out of a vast range of available datasets. With users interacting with search engines, sufficient information is already hidden in the log files. Compared with explicit feedback data,information that can be derived/extracted from log files is virtually free and substantially moretimely.In this dissertation,I propose adeep learning based ranking framework that can learng and update the ranking function based on user behavior data. The contributions of this framework include 1)a log processor thatthat can ingest, extract user access pattern and create training PREVIEW data from Web log in a batch mode/real-time;2) a query understanding module to better interpret users search intentusing weblog processing results and metadata; 3) a feature extractor that identifies ranking features representingusers search interestsof geospatial data; and 4) a deep learningbased rankingalgorithm thatautomatically learns and updates a ranking function based on user behavior.The search ranking resultswill be evaluated using precision at K and normalized discounted cumulative gain (NDCG).This research will strengthen ties between Earth observations and user communities by addressing the ranking challenge,the fundamental obstacle in geospatial data discovery.
Name of the Researcher:  Yongyao Jiang
Name of the Supervisor(s):  Dr. Chaowei Yang
Year of Completion:  2018
University:  George Mason University
Thesis Link:   Home Page Url