Abstract
Recommender systems are personalisation tools that predict a user’s preference for an item. They are used on services where users have to choose between a large number of options, such as shopping services like Amazon and movie websites such as Netflix. Even with search functions, the choice can often be overwhelming, and recommender systems provide users easy access to the items that are best suited to them, usually based on their history of previous preferences. Algorithmically, recommender systems often generate a score between 0 and 1 which represents how much a user will like an item. Items with high scores can then be recommended.Reciprocal recommender systems are a more complex subtype of recommender system designed for services where the objective is to recommend people to each other, such as online dating, social and recruitment services. They are considered complex because the recommendation must be based on a bidirectional preference relation: it is important that both the person being recommended and the person viewing the recommendations are satisfied.
In spite of the relatively interesting algorithmic challenge their complexity presents, reciprocal recommender systems have been overlooked in the literature, with a rich variety of research concentrating on user-item recommendation, and very few techniques for reciprocal recommendation. The purpose of this PhD is to contribute new methods and ideas to reciprocal recommendation, to advance the field with modern techniques currently being used in user-item recommendation, and to develop novel algorithms unique to reciprocal recommendation.
This thesis is divided into three main sections: content-based filtering, collaborative filtering and hybrid filtering, to correspond to the main subdivisions of recommender systems. Each of these sections contributes new techniques to that field within the context of reciprocal recommendation. This includes both adaptations of existing algorithms and entirely novel methods.
All of these methods are tested against large datasets from industry, including data from a popular online dating service, and from a social recipe-sharing website. Their success over and above the current state of the art demonstrates the value of these new techniques, and provides a base of modern techniques that can be further improved upon by researchers in this field.
Date of Award | 12 May 2022 |
---|---|
Original language | English |
Awarding Institution |
|
Supervisor | Ryan McConville (Supervisor) & Weiru Liu (Supervisor) |
Keywords
- Machine Learning