|
Title
|
|
|
|
Offline approaches to recommendation with online success
| |
|
Author
|
|
|
|
| |
|
Abstract
|
|
|
|
| Recommender systems are information retrieval applications that provide users with algorithmic recommendations, in order to assist decision-making when sufficient knowledge about the various options is lacking. These systems have known widespread adoption in recent years and are extensively used by digital platforms to suggest restaurants, books, musical artists, retail products and even romantic partners — much like the recommendations you could provide for a friend. Modern approaches to recommendation typically follow the "collaborative filtering" paradigm, making use of a large dataset of user behaviour to infer preferences, and to subsequently predict your preferences based on your historical behaviour. Impressive advances in machine learning in recent years have found their way to the recommendation field, and these systems are becoming ever more accurate when it comes to learning and predicting user preferences. There is, however, a gap between the recommendation use-case that is often posed in academic research and the use-case that practitioners typically face in industry. The work in this dissertation focuses on fundamental advances in three areas. First, the research literature typically deals with a single model trained on a static dataset. In contrast, recommendation models in the real world are often part of a dynamic ecosystem where new data is constantly coming in and models need to be kept up-to-date to remain competitive. In such settings, a clear need arises for models that can be computed efficiently and incrementally. Second, newly proposed methods in the research literature are often evaluated using offline procedures on datasets containing user-item interactions. While this is common practice in the broader machine learning field, recommendation datasets often lack true “labels”. As a result, currently existing evaluation procedures are notoriously uncorrelated with those obtained via the golden standard in industry of using online experiments such as randomised control trials — also known as A/B-tests. Third, most approaches to recommendation learn from observational datasets consisting of user-item interactions. As they do not take any information regarding previously shown recommendations and their outcomes into account — they take a purely passive stance that focuses on prediction, which contrasts with the interventionist nature of the problem in practice. Indeed, we wish to show recommendations to users in hopes of encouraging engagement and satisfaction. As such data is being generated by every online platform with a recommendation component — it comes naturally that the value of such datasets needs to be explored. |
| |
|
Language
|
|
|
|
English
| |
|
Publication
|
|
|
|
Antwerpen
:
Universiteit Antwerpen, Faculteit Wetenschappen, Departement Informatica
,
2021
| |
|
Volume/pages
|
|
|
|
x, 169 p.
| |
|
Note
|
|
|
|
:
Goethals, Bart [Supervisor]
| |
|
Full text (open access)
|
|
|
|
| |
|