| Author: | Liao, Shuiying |
| Title: | Data-driven recommendations for fashion − an investigation of personalization with sparse data |
| Advisors: | Mok, Tracy (SFT) |
| Degree: | Ph.D. |
| Year: | 2025 |
| Department: | School of Fashion and Textiles |
| Pages: | xv, 159 pages : color illustrations |
| Language: | English |
| Abstract: | The fashion industry is a dynamic and rapidly evolving field, characterized by diverse consumer preferences, behaviors, and styles. This ever-changing landscape has made innovative fashion recommendation systems a crucial area of study. In data-driven recommendation tasks, data sparsity issue is of significant concern, and it becomes even more critical in fashion recommendations, where a high level of personalization is required to cater to individual preferences. However, this emphasis on personalization exacerbates the data sparsity challenge, as user-specific interactions are often limited. While existing data-driven recommendation methods have made significant strides in fashion recommendation research, they still face notable limitations. The most significant challenges arise from their inability to effectively address data sparsity, leading to suboptimal performance. This issue extends to their capacity to incorporate user-specific preferences effectively. Simple linear models, which are often employed in personalization tasks, fail to capture the complicated relationship between user preferences and item compatibility. Other limitations include challenges in tracking dynamic user preferences, understanding the multifaceted nature of fashion items, and providing context-aware recommendations that reflect the user’s specific needs at any time. Capturing both enduring preferences and temporary needs is crucial for personalized recommendations. Long-term interests, such as style or brand affinity, and short-term interests, influenced by trends or seasons, both play a role in creating relevant suggestions. Ultimately, these challenges are fundamentally rooted in data sparsity. This thesis addresses the above limitations by proposing multiple strategies specifically designed to manage sparse data in the context of data-driven fashion recommendations. By focusing on two core tasks—personalized clothing matching and sequential recommendation—the study develops methods to better utilize sparse data and improve personalization and recommendation outcomes. Specifically, three alternative methods are designed for the personalized complementary recommendation task from three perspectives—consistency, coupled indirect personal compatibility, similar users or products—and can be flexibly replaced or combined depending on specific application requirements, and one method are proposed for the second sequential recommendation task. In Chapter 3, innovative constraints based on consistent user behaviors are introduced to better utilize interaction history, enhancing the model’s ability to infer preferences with limited data. Chapter 4 presents a novel concept, Indirect Personal Compatibility, to balance personalization and compatibility in recommendations, achieving alignment with individual preferences through iterative training. Chapter 5 explores contrastive learning to capture latent representations in sparse environments, from user preferences and item compatibility views, and implements adaptive collaborative signal selectors to mitigate data noise when using historical interactions as auxiliary information. Last but not least, in product sequential recommendation in Chapter 6, besides a multi-scale transformer architecture, which captures both long-term and short-term preference within a unified framework. Data sparsity issue is addressed by identifying similarities among items the user has interacted with, filling the sparse interaction matrix across various behaviors. Besides, multiple data mining and augmentation techniques are explored to expand upon the available data, uncovering new chances for enhancing recommendation accuracy in sparse environments. Extensive experiments across two tasks and multiple open-access datasets to evaluate the effectiveness of all proposed methods in tackling data sparsity in data-driven personalized fashion recommendations. Results show significant improvements in recommendation performance, underscoring the potential of these approaches to enhance both robustness and accuracy in data-sparse environments. |
| Rights: | All rights reserved |
| Access: | open access |
Copyright Undertaking
As a bona fide Library user, I declare that:
- I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
- I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
- I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.
By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.
Please use this identifier to cite or link to this item:
https://theses.lib.polyu.edu.hk/handle/200/14305