| Summary: | This research explores the technical feasibility of developing a knowledge-based apparel style recommendation system through investigations on apparel communication theory, data construction and machine learning techniques. It intends to improve the poor user experiences of online clothes shopping caused by the unpractical style searching, recommendation and personal styling engines. This study started with building up the theoretical foundation of apparel data and recommendation system. Then, an apparel data coding method and two apparel datasets are developed based on the apparel communication system and semiotics theory. ATTRIBUTE dataset captures natural and design features while MEANING dataset labels communicative meanings on style and body. Thirdly, the technical feasibility is investigated by statistics analytical methods to evaluate data relations and machine-learning methods to learn from the training data and predict apparel MEANINGs. The author found that the proposed data might exist non-linear relations, which restricts statistics analytical methods. Instead, machine-learning based methods are applicable as evidenced by three apparel MEANING prediction models. The three models also prove that the new apparel data coding method and ATTRIBUTE dataset could enhance the learning model since it captures more accurate apparel features. Additionally, the most useful data learning method is identified when it firstly learns ATTRIBUTEs from images via CNN model, and then determines MEANINGs from predicted ATTRIBUTEs by LKF classifier. The conclusion from this research is that it is technically feasible to develop an apparel style recommendation system. This research contributes a new method to the field of apparel recommendation system study. It fills the gap of lacking deep understandings of apparel knowledge. The proposed approach made three improvements: (1) a profound theory of apparel as a foundation, (2) a new apparel dataset construction method capturing design features and connotative meanings, and (3) the image-attribute collaborated data training model, which can effectively recognise in-depth design features and make precise predictions on connotative meanings.
|
|---|
