Thesis
The users’ behaviour in audio streaming services : investigations in the music and podcast domains
- Creator
- Rights statement
- Awarding institution
- University of Strathclyde
- Date of award
- 2024
- Thesis identifier
- T16912
- Person Identifier (Local)
- 201979732
- Qualification Level
- Qualification Name
- Department, School or Faculty
- Abstract
- In recent years, online audio streaming services (e.g., Amazon Music and Spotify) have witnessed an increase in popularity due to content digitisation. These platforms, now offering on-demand music, personalised playlists, and recently, podcasts, have significantly transformed users’ behaviour and their interactions with these platforms [1]. Podcasts, defined as spoken documents that can be represented by their transcriptions [2, 3], are swiftly becoming a central medium for online information seeking activities. Due to their considerable demand, streaming services have expanded their catalogues to include podcasts alongside music [4, 5]. Thus, there is an important research need for effective, cross-domain, and multi-modal, information access tools and methods that can guide users through these vast content libraries by aligning with their preferences and needs. However, despite the research relevance, understanding, modelling, and predicting how users interact with content on such streaming services remains under-researched [3, 6]. This thesis aims to address this gap by delving into the nuances of users’ behaviour in order to improve our overall understanding. This is motivated by the invaluable stream of information that an accurate representation of the users’ behaviour can provide to the underlying recommendation process. In particular, the focus of this thesis is the intricate relationship between understanding users’ behaviours, predicting these, and developing novel user-centric interfaces that are informed by these findings. This research is performed across both the music and podcast domains, aiming to unravel new facets of users’ behaviour and thus inform novel user modelling and recommendation techniques. The first part of the thesis focuses on understanding and predicting users’ behaviour in the music domain. In particular, it presents extensive investigations into users’ skipping behavior during listening sessions. Chapter 3 introduces a novel approach to identify fine grained session skipping behaviors. Four major session skipping patterns are identified through extensive evaluation, namely the listener, listen-then-skip, skip-then-listen, and skipper. A subsequent analysis of the differences among these patterns under varying listening contexts is also presented. With a deeper understanding of the users’ music skipping behaviour, Chapter 4 investigates the utility of users’ historical data for the task of sequentially predicting users’ skipping behaviour. To this end, the applicability and effectiveness of Deep Reinforcement Learning (DRL) for this task is demonstrated. An in depth post-hoc and ablation analysis indicates that users’ behaviour features are the most discriminative of how the proposed DRL model predicts music skips. Content and contextual features are reported to have a lesser effect. The second part of the thesis delves into the podcast domain. Chapter 5 introduces Podify, the first web-based podcast streaming platform specifically designed for academic research. Resembling existing streaming services, Podify supports academic research in the podcast domain, specifically in the under-researched areas of search and user behavioural analysis. Chapter 6 and Chapter 7 present, respectively, the methodology of a user study conducted through Podify and a discussion on the impact of text-based components, such as captions and full-text transcriptions, on how users assess the relevance of podcast content to their information needs. This is motivated by their well-established multidimensional role for improved information accessibility [7,8], and the alignment with the principles of Universal Design [9, 10], which support the ability to cater to diverse audiences and learning styles [11,12]. By combining qualitative (i.e., the participants’ reported relevance judgements of podcasts) and quantitative (i.e., listening activity) data, the importance of these textual components in enabling users to better assess the relevance of podcast content is shown.
- Advisor / supervisor
- Revie, Crawford
- Levine, John
- Moshfeghi, Yashar
- Resource Type
- DOI
Las relaciones
Elementos
| Miniatura | Título | Fecha de subida | Visibilidad | Acciones |
|---|---|---|---|---|
|
|
PDF of thesis T16912 | 2024-05-08 | Público | Descargar |