User Embedding from Ranking

Description: Previous work showed that we can reconstruct a face someone is thinking of only using their ranking of multiple faces based on the similarity with their mental face. Training such systems requires large amounts of human ranking data which would be time consuming and expensive to collect. Instead, simulating human ranking behaviour allows us to generate arbitrary amounts of training data. However, previous work showed that the correlation of face rankings between humans is low. Therefore, the goal of this project is to create user specific embeddings from their ranking data, which can then be used to simulate user specific face rankings. Your task will be to develop a method that allows us to calculate user embeddings from ranking data and how these can be integrated into a computational user model. To train and evaluate your method, you will be using an existing dataset of humans ranking faces and extend it if necessary.

Supervisor: Florian Strohm

Distribution: 45% Implementation, 10% Data Processing, 45% Evaluation and Analysis

Requirements: Good Python knowledge, ideally experience with deep learning, interest in user modelling