| Author: | Ding, Yujuan |
| Title: | Deep learning-based fashion advising |
| Advisors: | Wong, Calvin (ITC) |
| Degree: | Ph.D. |
| Year: | 2021 |
| Subject: | Clothing trade -- Data processing Fashion merchandising -- Data processing Human-computer interaction Machine learning Recommender systems (Information filtering) Hong Kong Polytechnic University -- Dissertations |
| Department: | Institute of Textiles and Clothing |
| Pages: | xiv, 124, [19] pages : color illustrations |
| Language: | English |
| Abstract: | Fashion advising provides proper fashion suggestions to facilitate the decision-making processes, which is helpful for both ordinary people and fashion business. In the recent years, with the accumulation of fashion-related data and the development of deep learning technologies, data-driven fashion analysis based on deep learning has attracted great research attention. This thesis focuses on the deep learning-based fashion advising for ordinary users/individuals, to which personalization and fashionability are two key perspectives. The two perspectives correspond to two basic standards in making desired fashion suggestions, to cater to the user preference/taste in fashion, to offer fashionable guidance and improve their aesthetic ability in fashion. In accordance with personalization and fashionability in fashion advising, this thesis works on two specifc tasks which are personalized fashion recommendation and fashion trend foresting. Existing works on personalized fashion recommendation limited to leverage more characteristic attributes or enhance the visual information of fashion items. However, the shopping patterns in user behaviors and the short-term behavior transition in fashion shopping were overlooked in developing the recommender systems. In the feld of fashion trend forecasting, the research of data-driven fashion forecasting is still at its early stage. On the one hand, the fashion elements focused in previous studies are not specifc, fine-grained and comprehensive to reveal real fashion trends. On the other hand, the proposed models were still based on statistical models which fall shorts in handling complicated fashion trend signals. To address the limitations in existing studies, this thesis proposes three main objectives and fulflls them with three deep learning-based approaches accordingly. First, to effectively capture the user preference in fashion and thus facilitate the performance of fashion recommendation, this thesis proposes to model the underlying shopping patterns in fashion shopping behaviors, which imply the diverse user preferences under specifc aspects, such as style, brand or print pattern. For such goal, a Field-aware Graph Collaborative Filtering (FGCF) method is proposed to capture the fined-grained user shopping patterns which have been widely ignored in previous research. Specifcally, the proposed FGCF method is able to model the factor feld-level interactions and make overall recommendation prediction by aggregating the feld-level results. It not only predicts the holistic user-item preference, but also infers the specifc fashion preferences in different factor felds. Extensive experiments on real-life fashion purchase data demonstrate the effectiveness of the proposed FGCF method. Second, to effectively model the short-term transition of user behaviors in fashion shopping and develop better sequential fashion recommender system, this thesis proposes an Attentional Content-level Translation-based Fashion Recommender (ACTR). Specifcally, the ACTR leverages the item-item relationships (indicating the short-term intentions) in modeling the item-item interactions. To tackle the sparsity problem in item-item interactions, it introduces the content-level item transition modeling which decomposes the overall item-item interaction into different fashion aspects. Moreover, a user-aware content attention mechanism is devised in the ACTR to properly aggregate the content-level modeling results and generate the final recommendation results. Extensive experiments on real-life fashion shopping data demonstrate the effectiveness of the proposed ACTR method. Third, towards meaningful fashion trend forecasting, this thesis aims to analyze fine-grained fashion elements which can effectively reveal fashion trends, in specifc, to model and forecast the fashion trend of specifc fashion elements for specifc user groups. To this end, a large-scale fashion trend dataset (FIT) is firstly collected from Instagram and the time series popularity records of fashion elements as well as user information are extracted. To effectively model the time series data of fashion elements with rather complex patterns, a Relation Enhanced Attention Recurrent (REAR) network is proposed, which takes advantage of the capability of deep recurrent neural networks in modeling time-series data and connects specific fashion trends through the relations between user groups and fashion elements. Extensive experiments have demonstrated that REAR can make solid and meaningful fashion trend forecasting for a period of time in the future. In summary, this thesis works on the deep learning-based fashion advising problem from two different key perspectives and studies three specific sub-tasks. With the output of the three studies, the personalized and fashionable fashion advice are able to be generated for specific users based on the specific approaches proposed. The research problems extracted and addressed in this thesis promote the development of personalized fashion recommendation and fashion trend forecasting. Moreover, the output of the thesis has a strong impact on the entire fashion industry which can specifically benefit the process of design, manufacturing and retailing. |
| 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/200/11139