|
Title
|
|
|
|
Temporal dynamics in online recommender systems
| |
|
Author
|
|
|
|
| |
|
Abstract
|
|
|
|
| In this thesis, we describe several methods to handle temporal dynamics in recommender systems. Recommender systems are often deployed in highly dynamic environments such as news websites or retail pages. When a new article is published, it is important that the recommender system is able to quickly recommend it, otherwise, the users will fail to find utility in the recommendations. Utilising temporal information proves to be a crucial factor to make sure these systems perform well and generate the right recommendations for users browsing news or retail websites. First, we look at how to keep models up to date, such that they can deal with the constantly changing online environments, and recommend new items quickly. Next, we show that using only more recent data or giving more recent data additional weight, makes baseline algorithms perform impressively well. This finding, going against the common knowledge that more data is better, indicates that the quality of data is a fundamental part of making good recommendations. Finally, we investigate the effects popularity has on recommendation quality, and find that a balance needs to be found between popularity and relevance in order to give the best possible recommendations. |
| |
|
Language
|
|
|
|
English
| |
|
Publication
|
|
|
|
Antwerp
:
University of Antwerp, Faculty of Science
,
2023
| |
|
Volume/pages
|
|
|
|
vi, 118 p.
| |
|
Note
|
|
|
|
:
Goethals, B. [Supervisor]
| |
|
Full text (open access)
|
|
|
|
| |
|