Publication Type

PhD Dissertation

Version

publishedVersion

Publication Date

9-2019

Abstract

Personalized recommendation, whose objective is to generate a limited list of items (e.g., products on Amazon, movies on Netflix, or pins on Pinterest, etc.) for each user, has gained extensive attention from both researchers and practitioners in the last decade. The necessity of personalized recommendation is driven by the explosion of available options online, which makes it difficult, if not downright impossible, for each user to investigate every option. Product and service providers rely on recommendation algorithms to identify manageable number of the most likely or preferred options to be presented to each user. Also, due to the limited screen estate of computing devices, this manageable number maybe relatively small, yet the selection of items to be recommended is personalized to each individual users.

The basic entities of a personalized recommendation system are items and users. Personalization can be achieved through custom alternatives for delivering the right experience to the right user at the right time on the right device. Therefore, personalized recommendation can appear in many forms, depending on the characteristics of the items and the desired experience that the system wants users to have. In this thesis, we encompass two perspectives on personalized recommendation: preference learning and similarity learning. The former refers to the personalization in which the recommendation is tailored towards users' preference. The latter, on the other hand, refers to personalization approach in which recommendation is generated based on the users' personal perceptions of similarity between the items.

In the preference learning perspective, we focus on the task of retrieving recommendations efficiently and propose two techniques for this objective. For the first technique, we rely on Euclidean embedding to learn user and item latent vectors from users' ordinal preferences. Since they operate in the Euclidean space, these latent vectors natively support efficient nearest neighbor search using geometric structures such as spatial trees. For the second technique, our key idea is to desensitize the effect of vector magnitudes when modelling users' preferences over items. That effectively reduces the recommendation retrieval problem to the nearest neighbor search problem with cosine similarity, which can be solved efficiently with various indexing methods such as locality sensitive hashing, spatial trees, or inverted index. Extensive experiments on publicly available datasets show significant improvement of proposed techniques over the baselines.

In the similarity learning perspective, we are interested in the setting where there are multiple similarity perceptions in the data. Towards modelling these perceptions effectively, we propose two approaches that are natively multiperspective. One is a graph-theoretic framework that yields a similarity measure for any pair of objects for a perspective. Another is a geometric framework that learns multiple low-dimensional representation of objects, each for one perspective. Experiments in both studies show that the adoption of multiperspective approach allows us to better model the similarity between objects, as compared to classical uniperspective methods, which ignore the multiperspectivity in the data.

Keywords

Personalized Recommendation, Indexing, Recommendation Retrieval, Similarity Learning, Multiperspective Similarity Learning

Degree Awarded

PhD in Information Systems

Discipline

Computer Engineering | Programming Languages and Compilers

Supervisor(s)

LAUW, Hady Wirawan

Publisher

Singapore Management University

City or Country

Singapore

Copyright Owner and License

Author

Download

Share

COinS