Computational Fashion Preference Modeling

Abstract:

Human fashion preference expresses each individual's personalities by adopting or composing a set of preferable fashion elements. Properly understanding and capturing human fashion preference provide important value to the fashion industry. However, traditional research on fashion preference is mainly carried out by fashion experts, which is highly professional and inefficient as it requires a lot of manually crafted data. With the fast development of the Internet, Data Science, and Artificial Intelligence in recent decades, novel computational solutions are surfacing to revamp this traditional research area, and we term this strand of research as Computational Fashion Preference Modeling. In this thesis, we progressively introduce three attempts in computational fashion preference modeling as detailed below.

First, we aim to extract human-centered occasion-aware fashion knowledge from social media. Human has accumulated a large amount of general knowledge about what to wear and how to wear, which should not only address physiological needs but also the requirements in social events and activities. In this study, we propose a novel method to automatically harvest such fashion knowledge from unstructured social media data. Even though the extracted fashion knowledge contains rich information, it is still limited to capturing the temporal dynamics and location-aware variance of fashion preference.

Second, to address the limitations above, we study the problem of fashion trend forecasting based on social media. Fashion trend forecasting is a crucial task for both academia and industry. Although some efforts have been devoted to tackling this challenging task, they only studied limited fashion elements with highly seasonal or simple patterns, which could hardly reveal the real fashion trends. Towards insightful fashion trend forecasting, this work focuses on investigating fine-grained fashion element trends for specific user groups. In addition, we seek to incorporate rich domain knowledge to enhance forecasting performance. Despite the insights derived from fashion trend forecasting, the granularity of cities and fashion elements is still insufficient when directly applied to real applications such as recommendations.

Third, in view of personalized fashion preference modeling, we tackle the problem of sequential fashion recommendation in E-commerce platforms. To capture the two types of key patterns, i.e., the user-item interactions and item-item transitions, and address the data sparsity problem, we propose to leverage two types of global graph, i.e., the user-item interaction graph and item-item transition graph. Despite the promising performance in the item-level recommendation, it is incapable to model people's preference over outfits, which is composed of a set of compatible items.

Finally, to tackle the problem of personalized outfit recommendation, we are confronted with two types of challenges: 1) more complicated patterns, including user-outfit interaction, user-item interaction, and outfit-item affiliation; and 2) sparse and noisy interactions. To address these challenges, we first propose a light but effective backbone model which can capture the three types of patterns. Second, we introduce contrastive learning as an auxiliary task to deal with the problem of sparse and noisy interactions.