Interactive learning for multimedia archive exploration

Interactive learning for multimedia archive exploration

Hammudoğlu, Joren (TU Delft Electrical Engineering, Mathematics and Computer Science)

Liem, Cynthia (mentor)
Scharenborg, Odette (graduation committee)
Yorke-Smith, Neil (graduation committee)

Delft University of Technology

2019-08-27

Recommender systems are essential for filtering immense amounts of available digital content. As these quantities keep increasing, the impact of recommendations does so as well. In this work, we address negative impacts current state-of-the-art recommenders have. For the algorithmic filtering of items that are recommended to users, collaborative filtering techniques have been shown to give good performance in terms of accuracy. However, at the same time, they suffer from several drawbacks. Beyond the cold-start problem, which causes low recommendation quality for unpopular or newly added items and users, they also risk the creation of filter bubbles. Moreover, they tend to homogenise user preference and they recommend items with low content diversity.
We argue that interactive recommenders can alleviate the problems because they, unlike the state-of-the-art, learn from all user interactions. We propose a method for combining interactive recommenders with the explorative capabilities of multi-armed bandit algorithms, which we use to formulate three explorative content-based recommenders. Using a simulation method, we evaluate them by analysing their accuracy, the diversity of recommendations and user preference, and the change in the distributions of preferences. The results show an increased content diversity, less homogenisation of preferences, and lower susceptibility to the formation of filter bubbles.

Recommender System
Interactive Learning
Multimedia retrieval

http://resolver.tudelft.nl/uuid:0c3beb54-6263-4739-ae37-c589147f1dff

Student theses

master thesis

© 2019 Joren Hammudoğlu