64 6. PERSONALIZED COMPATIBILITY MODELING
User 1
User 2
User 3
Figure 6.1: Examples of users’ outfit compositions.
may deliver the user preferred item brand or fabric. erefore, how to fully take advantage of the
multi-modal data in the context of the personalized clothing matching is a crucial challenge.
To address the aforementioned challenges, we present a personalized compatibility mod-
eling scheme for clothing matching, named GP-BPR, as shown in Figure 6.2, which is able to
measure the compatibility between fashion items from not only the general aesthetics but also
the personal preference perspectives. In particular, GP-BPR consists of two essential compo-
nents: general compatibility modeling and personal preference modeling. e content-based general
compatibility modeling works on learning the latent compatibility space shared by complemen-
tary items to characterize the item-item interactions toward clothing matching. Meanwhile,
the personal preference modeling focuses on exploiting the latent preference factor based on the
multi-modal data of fashion items and hence captures the user-item interactions comprehen-
sively. Ultimately, based on the BPR framework [100], GP-BPR jointly integrates the general
compatibility and personal preference modeling. To facilitate the evaluation, we construct a
large-scale dataset from the online fashion community IQON,
1
which comprises 308,747 out-
fits created by 3,568 users with 672,335 fashion items.
6.2 RELATED WORK
Owing to the recent booming of the fashion industry, increasing research attentions from both
the computer vision and multimedia communities have been paid to the fashion domain, espe-
1
https://www.iqon.jp/
6.2. RELATED WORK 65
Tops > Knit > Geometric
Wool Blend Sweater
Pants > Denim > Trousers
>Tomcat Jeans
Top
Top Latent
Embedding
User Historical Preference
Bottom Factor
Bottom Latent
Embedding
Bottom Latent
Embedding
Bottom Latent
Embedding
User Factor
User-visual
Factor
User-contextual
Factor
Bottom
top
knit
geometric
wool
blend
sweater
pants
denim
trousers
tomcat
jeans
Conv
TextCNN
Conv
TextCNN
General
Compatibility
Personal
Preference
BPR
Figure 6.2: Illustration of the proposed scheme. e general compatibility modeling aims to learn
the visual and contextual latent embedding of the items. e personal compatibility modeling
focuses on exploiting the latent user-item interaction factors to capture the user preference.
ese two components are integrated by the BPR framework to jointly tackle the personalized
clothing matching problem.
cially the clothing matching problem [24, 28, 31, 67, 108], which is usually cast as the compati-
bility modeling task between complementary fashion items. For example, Li et al. [67] proposed
an outfit quality predictor with the multi-modal multi-instance deep learning based on item ap-
pearance. In addition, Song et al. [108] introduced a content-based neural scheme toward the
compatibility modeling between fashion items based on their multi-modal data. Later, Yang
et al. [127] presented a translation-based neural fashion compatibility modeling framework,
which jointly optimizes the fashion item embeddings and category-specific complementary re-
lations in an end-to-end manner. Moreover, noticing that the fashion domain has accumulated
various valuable knowledge that can be helpful to guide the compatibility modeling, Song et
al. [109] shed light on integrating the rich fashion domain knowledge to the pure data-driven
learning, where a neural compatibility modeling scheme with attentive knowledge distillation
was presented. Although existing efforts have achieved compelling success, they mainly focused
on modeling the compatibility between fashion items purely based on the general item-item
compatibility and overlooked the user factor in the compatibility modeling, which is the major
concern of our work.
In addition, personalized recommendation in fashion domain also gains great research
attentions [9, 36, 114]. In particular, existing personalized recommendation work in fashion
..................Content has been hidden....................
You can't read the all page of ebook, please click here login for view all page.