Learning to Rank Places for Geospatial Search

I recently completed a master’s degree in data science at the Beuth University of Applied Sciences. As part of the study program, I authored a thesis titled Learning to Rank Places. Here are links to the thesis and presentation slides for the defence. Also, you can read the abstract below.

Ranking problems are omnipresent in interactions with software that retrieve information. Advancements in machine learning (ML) have led to novel solutions for solving ranking problems using a set of approaches known as Learning to Rank (LTR). The goal of this thesis is to demonstrate the effectiveness of learning to rank in solving the problem of ranking geographic places intended for navigation by comparing it to an existing place search engine called Pelias. Clickthrough logs collected from Pelias usage were utilized to create a training dataset for the learning to rank models. Linear, tree-based, and neural learning to rank models were built using the standard ML workflow and evaluated offline using the Mean Reciprocal Rank (MRR) metric. The tree-based models show significant MRR improvements over Pelias, while a subset of the linear and neural models show marginal improvements. An analysis of the results revealed open questions and clear directions for future work on the LTR models.