5.2. RELATED WORK 47
t
i
t
i
b
j
b
j
b
k
b
k
Attribute Classification
Networks
Attribute Representations of
Top-bottom Pairs
Latent
Prototypes
Compatible
Incompatible
Semantic Attribute
Representation
Attribute Classification
Networks
NMF
NMF
MLP
BPR
MLP
Sweater
Jeans
Pants
Figure 5.2: Illustration of the proposed scheme. We obtain the semantic attribute representa-
tions via the pre-trained attribute classification network, based on which we employ the NMF
framework to explore the latent compatible and incompatible prototypes. We jointly regularize
the latent prototype learning and compatibility modeling with the BPR framework.
regarded as the templates to guide the discordant attribute interpretation and the alternative
item suggestion. On the other hand, toward compatibility modeling, the proposed scheme seeks
the latent space to accurately measure the compatibility between fashion items using the MLP.
Ultimately, the proposed scheme seamlessly integrates the latent prototype learning and com-
patibility modeling with the BPR framework [100], where the pairwise preferences between
attribute prototypes and fashion items can be adaptively coupled and well exploited.
5.2 RELATED WORK
5.2.1 INTERPRETABLE COMPATIBILITY MODELING
As mentioned in Chapters 3 and 4, although existing studies [31, 67, 69, 86, 108] have achieved
compelling success in fashion analysis, especially the compatibility modeling, they mainly fo-
cused on utilizing deep learning methods to represent fashion items with the blurry semantic
features, resulting in their poor interpretability. To enhance the model interpretability, Feng
et. al. [24] proposed a partition embedding network to learn the embedding of each attribute
and then model the attribute-level compatibility between fashion items. Despite the promising
performance it accomplished, the attributes regarding the compatibility of fashion items can
be numerous yet they only adopted limited ones, making the interpretation incomprehensive.
..................Content has been hidden....................
You can't read the all page of ebook, please click here login for view all page.