Applications of Bayesian Recommender Systems in Online Environments

Abstract

Recommender systems play a vital role in how any well functioning website presents content to users. With the amount of content growing day by day, the ability to accurately hone in on the specific interests of any given user has taken on an ever increasing significance. In this thesis we investigate the application of a handful of techniques that are aimed at improving how content is presented to users in online environments. Two major real-world use cases are discussed: the recommendation of adverts on AutoTrader and the recommendation of news articles on the BBC website. Throughout the thesis emphasis is placed on the use and development of algorithms which scale to these real world use cases. The structure is a follows: Chapters 2 and 3 set out the groundwork for the major advances in this area and ensure that the knowledge required for topics in later chapters are available for the reader; these are predominantly focused on the multi-armed bandit problem and the variation inference algorithm. Chapter 4 describes the development of a new hybrid algorithm, FAB-COST, which seeks to leverage two moment matching variational inference algorithms, namely Expectation Propagation and Assumed Density Filtering, to build a procedure which is both accurate and able to handle steaming datasets. Its superior performance is demonstrated by benchmarking it against another state of the art contextual bandit algorithm using real user click data from the Auto Trader website. Chapters 5 and 6 demonstrate how recommendations can be improved at BBC News via Collaborative Topic Models. The fact that users have a strong preference towards recently published news, as well as the data pipeline the BBC have in place, means that cold-start presents a serious problem when making news recommendations. By incorporating contextual information present in the articles into the recommender system, articles that have no user histories associated with them can be recommended to users -- a task that traditional collaborative filtering algorithms are unable to fulfil. We give an in-depth mathematical and empirical analysis of the algorithms, as well as a detailed account of how they have been and can be used in a real world setting. We provide codes which demonstrate that the Collaborative Topic Model provides superior recommendations qualitatively and quantitatively, and importantly, works with the current infrastructure the BBC have in place.

Date of Award	31 Dec 2021
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Thomas House (Supervisor) & Neil Walton (Supervisor)