Naive Bayesian Network for Automated, Fashion Personal Stylist

N. Strain and J. I. Olszewska

School of Computing and Engineering, University of West Scotland, U.K.

Keywords:

Intelligent Systems, Agent-oriented Software Engineering, Expert System, Machine Learning, Bayesian

Networks, Knowledge-based System, Web Intelligence, Data Mining, Industrial Applications of AI.

Abstract:

Fashion is an area that people experience every day. Fashion can be seen as homogenizing, since encouraging

everyone to dress in a certain way that is inﬂuenced, e.g. by celebrities and social media. However, nowadays,

fashion is also a search for individuality and personal expression. Hence, this work is about the development

of an intelligent web application to help people by providing them with clothing suggestions based on their

previous garment selections at the registration stage and after determining each user’s individual style thanks

to machine learning techniques such as Naive Bayesian Networks. The resulting intelligent system has been

thoroughly tested on real-world datasets as well as successfully released to end-users.

1 INTRODUCTION

Fashion in clothing is a major economic force relying

on the design, tailoring, and dissemination of clothes

as well as their collections’ images, while impacting

the social and cultural life and thriving on novelty and

change.

Fashion is thought to have started during the Re-

naissance (Arnold, 2009), with the ﬁrst fashion mag-

azines appearing in the later 17th century and increas-

ing fashion visibility and desirability. In the 18th cen-

tury, a shift occurred from annual changes in textile

designs and fashion styles to seasonal changes. By

the end of the 19th century, fashion’s growth as a driv-

ing force within the clothing industry brought stylish

clothes to a wider cross-section of people. During the

20th century, ‘fast fashion’ came to supersede the in-

dustry’s previous seasonal timetable with regular sup-

plies of new garments sent out to high-street retailers,

while fashion consumers have increasingly sought to

individualize their look by customizing apparels and

mixing designer, high street, and vintage clothes. This

has enabled people to act as designers themselves, if

not of individual garments, then of the look and image

they wish to convey (Arnold, 2009).

The beginning of 21st century has been marked

by the development of the world-wide e-commerce

fashion industry, which is currently worth $545 bil-

lion and is expected to grow to $712.9 billion by 2022

(Started, 2019). Online fashion shopping represents

67% of UK online purchases, and this ﬁgure contin-

ues to rise (Started, 2019). For years, leading brands

such as Chanel were reluctant to have an online pres-

ence, but in order to meet demands of the current

digital world, they had to adapt. Indeed, good cus-

tomer service is no longer what fashion brands need

to stay competitive, since the millions of online shop-

pers look for a ﬂuid, interactive shopping experience

(Cecilio, 2015). On the other hand, nowadays, fash-

ion is not solely depending on the latest issue of Vogue

to keep people up to date with the latest trends, as

this is now being powered more and more through so-

cial media and, in particular, Instagram (Jaradat et al.,

2018). Currently, 200 million users of Instagram fol-

low at least one fashion account which 45% of them

use to gain inspiration for fashion trends and present

looks (Started, 2019).

Thence, fashion is an intricate ﬁeld which, beyond

garments and their imagery catalogues, brings to-

gether couture designers’ visions, dressmakers’ craft

skills as well as mass-production short-cycle outputs,

fashion media’s trends, and individuals’ aesthetic sen-

sibilities. Although cloth prices still provide the most

obvious constraint, the emphasis is actually placed

upon consumers’ ability to put together an interesting

and individual outﬁt.

For this purpose, some individuals call upon fash-

ion personal stylists to help them to look their best by

curating clothing, makeup, and other aspects of per-

sonal style. To overcome economic or geographical

constraints, platforms such as Thread (Thread, 2019)

can be used to bridge stylists and consumers, but they

continue to mainly rely on human agents’ services.

Consequently, they can still be biased, e.g. by picking

814

Strain, N. and Olszewska, J.

Naive Bayesian Network for Automated, Fashion Personal Stylist.

DOI: 10.5220/0009123808140821

In Proceedings of the 12th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2020) - Volume 2, pages 814-821

ISBN: 978-989-758-395-7; ISSN: 2184-433X

Table 1: List of adopted clothing attributes and related labels.

Attribute Labels

Gender Male, Female

Item Bottom, Top, Dress

Occasion Sport, Casual, Smart

Style Blouse, Bodycon, Chinos, Hoodie, Jacket, Jeans, Joggies, Jumper, Leggings, Midi, PoloShirt,

Shorts, Shirt, Skater, Skirt, Smock, Sweatshirt, Tie, Top, Trousers, T-shirt, Tux, Twill, Wrap

Fit Muscle, Oversized, Petite, Skinny, Slim

Keyword Bardot, Cable, Check, Chunky, Cold Shoulder, Crop, Drape, Floral, Funnel Neck, Leopard,

Off Shoulder, Oversized, Padded, Pattern, Pencil, Ribbed, Stripe, Tapered, Teddy, Tie, Vneck

Brand Adidas, Asos, Ax Paris, Barbor, Bershka, Boo Hoo, Boss, Daisy Street, Diesel, Farah, Fred Perry,

French Connection, G-Star, Hollister, Homme, JDY, Lacoste, Lee Luke, Lee Malone, Lee Rider,

Levis, Milk, Miss Guided, Miss Selfridge, Monki, Napapijri, New Look, Nike, Nudie, Only,

Pier One, Pretty Little Thing, Puma, Ralph Lauren, Replay, River Island, Stradivarius, Ted Baker,

Tommy, Vero, Vila, Warehouse

Colour Black, Beige, Black, Blue, Brown, Burgundy, Green, Grey, Nude, Orange, Pink, Red, Stone

White, Yellow

Sleeve Sleeveless, Short, Long, 3/4 Length

Length Short, Mini, Midi, Long

Material Cotton, Denim, Leather, Polyamide, Polyester, Satin, Viscose, Wool

clothes the personal stylist likes and not necessary the

ones the user likes. On the other hand, these services

are not expandable at the rate that technologies based

on Artiﬁcial Intelligence (AI) can (Luce, 2019).

Indeed, AI techniques can offer a scalable solu-

tion to a wide number of customers, while provid-

ing each of them with a personalized experience and

bespoke advice (Luce, 2019). Hence, since the last

decade (Chen et al., 2013), AI systems have started to

be developed in the retail industry to ﬁnd trends and

patterns among data, especially for intelligent fashion

analysis (Saponaro et al., 2018) such as clothing mod-

eling (Kim and Cho, 1999), clothing recognition (Seo

and Shin, 2018), clothing parsing (Yamaguchi et al.,

2012), clothing retrieval (Rubio et al., 2017), clothing

pairing (Stan and Mocanu, 2019), or clothing sugges-

tion (Liu et al., 2014).

In particular, clothing recommendation, i.e. cloth-

ing pairing and/or clothing suggestion, can be per-

formed by applying machine learning methods such

as Neural Networks (NN) (Takagi et al., 2017), (Jara-

dat et al., 2018) or Support Vector Machines (SVM)

(Chen et al., 2013), (Liu et al., 2014), (Quinn and

Olszewska, 2019) as well as Self-Organizing Map

(SOM) (Qian and Dong, 2010), (Hao and Hao, 2019)

or association rules (Wakita et al., 2015). How-

ever, the resulting current fashion expert systems are

mainly focused at making automated decisions on ex-

isting clothes within a customer’s wardrobe, based

on criteria such as a speciﬁed occasion (Liu et al.,

2014), season (Hao and Hao, 2019), scenery (Jo et al.,

2019), geo-location (Abe et al., 2018), garment colour

(Tu and Dong, 2010) or brand (Wakita et al., 2015).

Moreover, these recommenders rely on large amount

of data that are not always available or could infringe

on data privacy rights (Luce, 2019).

Actually, AI stylists face a number of challenges

that makes fashion understanding complicated, such

as its high level of variability (e.g. through time, sea-

sons, climates, region, culture, etc.) (Takagi et al.,

2017) and subjectivity (e.g. because of age, ethnicity,

places visited, personal interests, mood, etc.) (Priyad-

harsun et al., 2018). A major task lies thus in the

appropriate choice of attributes which describe the

properties of the clothing (Zhao et al., 2017). In-

deed, attributes are used both to learn a user’s personal

style and to make recommendations about what outﬁt

this customer should wear. It is worth noting that at-

tributes in fashion could be visual (Inoue et al., 2017),

i.e. directly extracted from fashion images by means

of computer-vision methods (Olszewska, 2019), or

semantic ((Valle et al., 2018) with garments being of-

ten described by their materials, colors, patterns, ﬁt,

and cut (Al-Halah et al., 2017).

In this paper, we propose an AI-personal-stylist

web application that embeds a Naive Bayesian Net-

work (NBN) in order to determine suitable, new

clothing suggestions in an automatic way, being given

user’s previous selections that are labeled with se-

mantic features and taking into account customer’s

age and gender, which are provided by the user him-

self/herself at the registration stage of the application.

The adopted machine learning method, i.e. NBN,

is thus a supervised approach which requires only a

small amount of data (voluntarily provided by the

user) to train the system and which can intrinsically

cope with missing data. Ultimately, our approach al-

lows to enhance user’s wardrobe by automatizing the

personal-stylist recommendations of new outﬁts (i.e.

emulating the AI stylist), while keeping the human in

Naive Bayesian Network for Automated, Fashion Personal Stylist

815

the loop (i.e. implicating the user), since ‘The cus-

tomer is the ﬁnal ﬁlter. What survives the whole pro-

cess is what people wear.’ – Marc Jacobs.

Although the dominant ﬁeld of fashion design is

womenswear (Hsiao and Grauman, 2018), this work

considers equally womenswear and menswear. On

the other hand, the web application analyzed personal

tastes of consumers (Liu and Shen, 2018) within both

Western fashion catalogues and online images which

were captured in the wild.

The resulting expert system provides a web-based

solution, since internet has been proven to be impor-

tant for fashion trends (Kitaura and Washida, 2015).

Moreover, this intelligent fashion engine is an inter-

active tool, because human-computer interaction is a

key to a successful recommender system (Wang et al.,

2015) and because it is important to understand not

only how algorithms see style but also how different

individuals see style (Takagi et al., 2017).

Thus, the contributions of this paper are twofold.

On one hand, as far as we know, we present the ﬁrst

study using Naive Bayesian Network for new clothing

suggestion. On the other hand, we propose a novel

web-based, interactive, AI-personal-stylist agent pro-

viding automated women/men fashion recommenda-

tions.

The paper is structured as follows. Section 2 de-

scribes the Naive Bayesian Network approach embed-

ded in our web-based expert system for new clothing

suggestion. Experiments are described in Section 3,

while conclusions are drawn up in Section 4.

2 PROPOSED AI-Stylist

RECOMMENDER

The developed, AI-stylist expert system consists in

an intelligent computer software that imitates human

decision making, in this case the ones of a personal

stylist, thanks to the use of a Naive Bayesian Network

(NBN) as explained in Section 2.1. The resulting,

automated recommender, that integrates this NBN, is

presented in Section 2.2, with an emphasis on its de-

sign as well as its features such as web accessibility,

user interactivity, and data security.

2.1 Naive Bayesian Network

The Naive Bayes (NB) model (Tan et al., 2018) is a

probabilistic classiﬁer that can perform estimations,

while handle data uncertainty, i.e. the lack of exact

knowledge due to inexact, incomplete, or even im-

measurable data (Negnevitsky, 2011). Moreover, NB

has the ability to process small datasets, i.e. c. 30+

samples (Webb et al., 2005). These properties make

the NB a suitable method for a fashion recommender,

since fashion is highly variable and subjective. In par-

ticular, the use of NB theory in expert systems allows

the computer to calculate independent assumptions.

Actually, the NB model assumes that all attributes are

conditionally independent given the class attribute.

The fashion attributes as used in this work are pre-

sented in Table 1. Since each attribute is learned sep-

arately, it simpliﬁes the teaching (or training phase)

of the algorithm, which is based on the Bayes rule, as

follows:

P(H|E) =

P(E|H).P(H)

P(E)

, (1)

where P(H|E) is the posterior probability of hypothe-

sis H upon observing evidence E, i.e. the conditional

probability that event H occurs given that event E has

occurred; P(E|H) is the likelihood, i.e. the probabil-

ity that hypothesis H being true will result in evidence

E; P(H) is the class prior probability of hypothesis H

being true; and P(E) is the predictor prior probability

(Negnevitsky, 2011).

The Naive Bayesian Network (NBN) does not di-

rectly use results from the training phase, but applies

parameters calculated from the training data to es-

timate the highest probability of the tested data in

the expert system’s testing phase. Thence, NBN can

be used in the fashion recommender to calculate the

probability of the user liking clothing suggestions

based on his/her previous selections. Since the NB

treats each attribute separately, this implies it copes

well with missing data, e.g. in consequence of the

user not having inputted all his/her characteristics or

the user having provided only a partial (i.e. not ex-

haustive) list of his/her garment tastes, as it happens

in real-world applications.

2.2 Developed Expert System

The present intelligent fashion recommender has been

designed following the Dynamic Systems Develop-

ment Method (DSDM) (Sommerville, 2015), which is

an Agile approach consisting of four stages, namely,

(i) business study and requirements, (ii) prototype it-

eration, (iii) design and development iteration, and

(iv) implementation.

In particular, the prototyping of the expert system

involves the training and testing of the NBN algorithm

which has been described in Section 2.1.

The NBN training, as depicted in Fig. 1, re-

quires the user to select the clothing item(s) s/he likes

among a database which is displayed within the ap-

plication and which constitutes the training dataset.

It is worth noting that the application is designed to

ICAART 2020 - 12th International Conference on Agents and Artiﬁcial Intelligence

816

Figure 1: Overview of the training process of the Naive

Bayesian Network.

ensure that the user picks at least one clothing item to

feed the expert system with. Indeed, the NB algorithm

needs yes/no data to be able to calculate a likelihood.

Hence, the clothing selected by the user are the ‘Yes’

data, and the remaining clothing items in the training

dataset that were not selected are the ‘No’ data.

In this application, the training dataset contains

113 records in total. Both genders have 40 registration

images and a remaining 20 clothing items for females

and 13 for males for the algorithm to select from. This

results from the fact that women’s clothing has tops,

bottoms, and dresses in the app, whereas males only

had tops and bottoms. On the other hand, each im-

age of training data is labelled using the 11 attributes

as presented in Table 1. There are other ﬁelds in the

database such as cloth ‘Price’ which are kept in the

database, but are not consider to be directly part of a

user’s fashion style model. The characteristics of each

of the 11 attributes are collected from this database to

be used in the frequency tables.

After a batch of new images, or testing datasest,

is uploaded into the system, the NBN testing is per-

formed as illustrated in Fig. 2. The NB algorithm is

then run using the new added items and considering

the previous yes/no data.

3 EXPERIMENTS

The developed AI-stylist agent integrating the pro-

posed NBN approach into the web-based, interactive

expert system has been evaluated both quantitatively

and qualitatively in a series of experiments as de-

scribed in Sections 3.1-3.3. It is worth noting that

the NB algorithm has been extensively tested follow-

ing the black box and white box approaches to ensure

that the recommender was producing and outputting

accurate results (as mentioned e.g. in the experiment

1), before it was released to the users for the usability

testing (as reported in the experiment 2).

Figure 2: Overview of the testing process of the Naive

Bayesian Network.

3.1 Datasets and Equipment

For all these experiments, two different datasets were

used. Both of them comprise data reﬂecting diver-

sity and equity, in order to minimize biases. Each

dataset contains the 11 ﬁelds, namely, Gender, Item,

Occasion, Style, Fit, Keyword, Brand, Colour, Sleeve,

Length, and Material, as presented in Table 1. These

ﬁelds are used in the algorithm for analysing patterns

within the data to make prediction.

The ﬁrst dataset is the primary database, which

was used in the deployed application, and constitutes

the main clothing catalogue that is based on Asos

fashion e-commerce site, with Asos granting us the

permission to use this data for the purpose of this re-

search project. Hence, this dataset allowed us to apply

our AI-stylist agent in real-world conditions.

The second dataset is composed of fashion images

in the wild. This data was obtained from copyright-

free images which were retrieved from Google. This

database is not only important to further assess the

NBN performance, but it also allows to appraise our

expert system integration with different sites or other

applications in the future.

3.2 Experiment 1

The ﬁrst series of experiments dwelt on the quantita-

tive evaluation of the integration process of the entire

system as schematized in Fig. 3. Some details of one

of these tests are reported as below.

In this test, we assume at ﬁrst that the NBN

has been previously trained using sample-female cus-

tomer’s choices which were interactively provided

during the registration phase. The resulting Clothing

Catalogue Table is generated automatically and re-

ported in Table 2 along with the used attributes which

have been limited to 4 categories for this example. For

each of these attributes, the likelihood is separately

Naive Bayesian Network for Automated, Fashion Personal Stylist

817

calculated and recorded in a table, as exempliﬁed in

Table 3 for the attribute ‘Item’.

Once the system is updated with new images of re-

cently released fashion outﬁts as schematized in Fig.

3, the NBN algorithm is processed to identify for a

particular user what garment has the higher probabil-

ity of ‘Yes’ within this data batch. Thence, the NBN

algorithm is run for each of the image cloth in order

to make a prediction on what the outcome of this user

liking this speciﬁc clothing item (e.g. as described in

Table 4) is. Considering the formerly computed like-

lihoods (e.g. Table 3) and based on the NB equation

(Eq. 1), the posterior probability is computed for each

evidence’s value, as detailed below:

P(H|E = Yes) 'P(Item = Top|H = Yes)

×P(Occasion = Casual|H = Yes)

×P(Style = Jumper|H = Yes)

×P(Colour = Green|H = Yes)

×P(H = Yes)

(2)

P(H|E = No) 'P(Item = Top|H = No)

×P(Occasion = Casual|H = No)

×P(Style = Jumper|H = No)

×P(Colour = Green|H = No)

×P(H = No)

(3)

However, NB is not able to calculate any values= that

have 0, because it would in turn zero out the en-

tire partial probability. This is known as the zero-

frequency problem. A way to overcome it is to use

the technique known as Laplacian Smoothing (or 1-

up smoothing) (Han et al., 2012); the previous results

(from Eqs. 2-3) becoming as follows:

P(H|E = Yes) '

'0.0119,

(4)

P(H|E = No) '

'0.0007.

(5)

Hence, the outcome of the NB prediction is the one

with the highest posterior probability (in this case, the

decision is a ‘Yes’, since P(H|E = Yes) > P(H|E =

No)). It is worth noting that Math.Log has been used

Table 2: Example of Clothing Catalogue Table.

Item Occasion Style Colour Decision

Top Casual Tshirt Grey Yes

Bottom Casual Jeans Blue Yes

Bottom Smart Chinos Black Yes

Dress Casual Twill Green Yes

Dress Casual Jumper Nude Yes

Top Casual Jumper Pink Yes

Top Casual Jumper Nude Yes

Dress Casual Skater Grey Yes

Top Casual Tshirt White Yes

Dress Casual Bodycon Black Yes

Top Casual Tshirt White No

Bottom Sport Leggins Black No

Top Smart Tshirt Red No

Bottom Sport Leggins Pink No

Top Sport Tshirt Pink No

Dress Smart Wrap Blue No

Top Sport Tshirt White No

Dress Casual Wrap Red No

Top Casual Tshirt Red No

Top Smart Shirt White No

Table 3: Example of the Likelihood Table which has been

generated based on Table 2 for the attribute Item’.

Item Decision

Yes No

Top 4/10 6/10

Bottom 2/10 2/10

Dress 4/10 2/10

Table 4: Example of a new clothing item and its attribute

values.

Item Occasion Style Colour

Top Casual Jumper Green

when calculating the probabilities in order to prevent

numeric errors that can occur with small real numeri-

cal values.

Lastly, the ﬁnal output of the NBN processing, i.e.

the item with the highest posterior probability of ‘Yes’

within the batch, is suggested to the user by sending

her an email as displayed in Fig. 3.

3.3 Experiment 2

The second type of experiments was focused on user

testing. Indeed, the AI-agent must provide users with

both an accurate algorithm for the clothe recommen-

dations and a service that fashion customers would be

interested to use.

For this purpose, a focus group, which was made

of 12 volunteer people with mixed ages and different

ICAART 2020 - 12th International Conference on Agents and Artiﬁcial Intelligence

818

Figure 3: Example of integration test.

(a) (b) (c) (d) (e) (f)

Figure 4: Examples of user testing for menswear and womenswear users, respectively. In particular, pictures (a)-(b) show a

sample of a menswear user’s given data within his application’s account, and picture (c) is a sample of automatically suggested

cloth data for the corresponding menswear user, while pictures (d)-(e) display a sample of a womenswear user’s given data

within her application’s account, and picture (f) is a sample of automatically suggested cloth data for the corresponding

womenswear user.

(a) (b)

Figure 5: Excerpt of survey results.

genders (i.e. 7 females and 5 males), has been con-

tacted to perform the usability tests of our intelligent

fashion application.

The group members were asked to interact with

the recommender and, in particular at the registra-

tion phase, to select the images they liked from the

clothing database in order to automatically build each

user’s style based on his/her choices. Then, they were

asked to provide a comment about the suggested out-

ﬁt which was emailed to them by the expert system as

well as a feedback about their overall experience with

the AI-stylist application.

The information gathered from the focus group at

the registration phase was thus used as the training

data for the supervised machine learning algorithm.

After processing all of the focus group’s images using

Naive Bayesian Network for Automated, Fashion Personal Stylist

819

the ‘Image Name’, ‘User Gender’, ‘Cloth Price’, and

the fashion attributes (see Table 1), the decision was

made automatically by the AI-agent for each of the

user and sent by email to each of them, respectively.

It is worth noting that, if the users were satisﬁed with

the received clothing suggestion, the algorithm was

considered as correct.

Thence, during the tests, each of these users was

given instructions to take screenshots of the clothing

items s/he selected (e.g. Figs. 4(a)-(b) and Figs. 4(d)-

(e), respectively) and of the clothing item that was

suggested by the recommender (e.g. Fig. 4(c) and

Fig. 4(f), respectively). After the tests, all of the

users were requested to complete a survey to assess if

the application has passed the user testing and to rate

their user’s experience. The survey results have been

anonymised and quantiﬁed as shown in Fig. 5. Over-

all, it appeared that the feedback about the clothing

item recommended by the AI-agent was very positive

(Fig. 5(a)). Moreover, all the users liked the usage of

our intelligent fashion application (Fig. 5(b)).

4 CONCLUSIONS

In this paper, we proposed a new AI-stylist recom-

mender which consists in an information ﬁltering sys-

tem that seeks to predict the ‘rating’, ‘preference’,

or ‘relevancy’ that a user would give to an item not

yet considered, using a model built on the character-

istics of a provided set of this user’s clothing items.

Hence, our intelligent fashion expert system auto-

matically suggests similar or related garments for a

given customer by applying a Naive Bayesian Net-

work (NBN). This supervised machine learning tech-

nique was trained based on user’s inputs, while tested

on different sets of fashion imagery as well as online-

retrieved images and deployed on data directly fed by

current fashion e-catalogues. Hence, the NBN algo-

rithm has been used for the ﬁrst time in a clothing-

suggestion, interactive web application, and the re-

sulting, intelligent stylist agent has shown promising

results in helping people to cope with fashion new

trends and seasonal shifts.

REFERENCES

Abe, K., Minoguchi, M., Suzuki, T., Suzuki, T., Akimoto,

N., Qiu, Y., Suzuki, R., Iwata, K., Satoh, Y., and

Kataoka, H. (2018). Fashion culture database: Con-

struction of database for world-wide fashion analy-

sis. In IEEE International Conference on Control, Au-

tomation, Robotics, and Vision, pages 1721–1726.

Al-Halah, Z., Stiefelhagen, R., and Grauman, K. (2017).

Fashion forward: Forecasting visual style in fashion.

In IEEE International Conference on Computer Vision

(ICCV), pages 388–397.

Arnold, R. (2009). Fashion: A Very Short Introduction. Ox-

ford University Press.

Cecilio, D. (2015). E-commerce is changing the

fashion industry: it’s time to catch up. The

Guardian. [online]. https://www.theguardian.com/

small-business-network/2015/feb/06/.

Chen, Q., Wang, G., and Tan, C. L. (2013). Modeling fash-

ion. In IEEE International Conference on Multimedia

and Expo (ICME), pages 1–6.

Han, J., Kamber, M., and Rei, J. (2012). Data Mining: Con-

cepts and Techniques. Morgan Kaufmann, 3rd edition.

Hao, L. and Hao, M. (2019). Design of intelligent clothing

selection system based on neural network. In IEEE

Information Technology, Networking, Electronic and

Automation Control Conference, pages 1789–1792.

Hsiao, W.-L. and Grauman, K. (2018). Creating capsule

wardrobes from fashion images. In IEEE Interna-

tional Conference on Computer Vision and Pattern

Recognition (CVPR), pages 7161–7170.

Inoue, N., Simo-Serra, E., Yamasaki, T., and Ishikawa, H.

(2017). Multi-label fashion image classiﬁcation with

minimal human supervision. In IEEE International

Conference on Computer Vision Workshops (ICCV),

pages 2261–2267.

Jaradat, S., Dokoohaki, N., Hammar, K., Wara, U., and

Matskin, M. (2018). Dynamic CNN models for fash-

ion recommendation in Instagram. In IEEE Interna-

tional Conference on Parallel and Distributed Pro-

cessing with Applications, pages 1144–1151.

Jo, S.-Y., Jang, S.-H., Cho, H.-E., and Jeong, J.-W. (2019).

Scenery-based fashion recommendation with cross-

domain generative adversial networks. In IEEE In-

ternational Conference on Big Data and Smart Com-

puting, pages 1–4.

Kim, H.-S. and Cho, S.-B. (1999). Development of an IGA-

based fashion design aid system with domain speciﬁc

knowledge. In IEEE International Conference on Sys-

tems, Man, and Cybernetics (SMC), pages III.663–

III.668.

Kitaura, S. and Washida, Y. (2015). Are geeks driving out

the fashion industry? In IEEE International Confer-

ence on Management of Engineering and Technology,

pages 1224–1229.

Liu, S., Liu, L., and Yan, S. (2014). Fashion analysis: Cur-

rent techniques and future directions. IEEE MultiMe-

dia, 21(2):72–79.

Liu, Y. and Shen, Y. (2018). Personal tastes vs fashion

trends: Predicting rating based on visual appearances

and reviews. IEEE Access, 6:16655–16664.

Luce, L. (2019). Artiﬁcial Intelligence for Fashion: How AI

is Revolutionizing the Fashion Industry. Apress.

Negnevitsky, M. (2011). Artiﬁcial Intelligence: A Guide to

Intelligent Systems. Addison-Wesley, 3rd edition.

Olszewska, J. I. (2019). Designing transparent and au-

tonomous intelligent vision systems. In Interna-

tional Conference on Agents and Artiﬁcial Intelli-

gence (ICAART), pages 850–856.

ICAART 2020 - 12th International Conference on Agents and Artiﬁcial Intelligence

820

Priyadharsun, S., Lakshigan, S., Baheerathan, S., Raja-

sooriyar, S., Rajapaksha, U., and Harshanath, S.

(2018). Parade in the virtual dressing room. In IEEE

International Conference on Computer Science and

Education, pages 458–461.

Qian, S.-Q. and Dong, A.-H. (2010). Application of Ko-

honen self-organizing map in fashion parts classiﬁca-

tion. In WASE International Conference on Informa-

tion Engineering, pages 226–229.

Quinn, M. and Olszewska, J. I. (2019). British sign lan-

guage recognition in the wild based on multi-class

SVM. In IEEE International Conference on Computer

Science and Informatics, pages 81–86.

Rubio, A., Yu, L.-L., Simo-Serra, A., and Moreno-Noguer,

F. (2017). Multi-modal joint embedding for fashion

product retrieval. In IEEE International Conference

on Image Processing (ICIP), pages 400–404.

Saponaro, M., Gal, D. L., Gao, M., Guisiano, M., and

Maniere, I. C. (2018). Challenges and opportunities

of artiﬁcial intelligence in the fashion world. In IEEE

International Conference on Intelligent and Innova-

tive Computing Applications, pages 1–5.

Seo, Y. and Shin, K.-S. (2018). Image classiﬁcation of ﬁne-

grained fashion image based on style using pre-trained

convolutional neural network. In IEEE International

Conference on Big Data Analysis, pages 387–390.

Sommerville, I. (2015). Software Engineering. Pearson,

London, 10th edition.

Stan, C. and Mocanu, I. (2019). An intelligent personalized

fashion recommendation system. In IEEE Interna-

tional Conference on Control Systems and Computer

Science, pages 210–215.

Started, G. (2019). The State of the Ecommerce

Fashion Industry: Statistics, Trends and Strat-

egy. [online]. https://www.shopify.com/enterprise/

ecommerce-fashion-industry.

Takagi, M., Simo-Serra, E., Iizuka, S., and Ishikawa, H.

(2017). What makes a style: Experimental analysis of

fashion prediction. In IEEE International Conference

on Computer Vision Workshops (ICCV), pages 2247–

2253.

Tan, P. N., Steinbach, M., and Kumar, V. (2018). Introduc-

tion to Data Mining. Pearson, 3rd edition.

Thread (2019). Stylist Recommendation Web Platform.

[online]. https://www.thread.com.

Tu, Q. and Dong, L. (2010). An intelligent personalized

fashion recommendation system. In IEEE Interna-

tional Conference on Communications, Circuits and

Systems, pages 479–485.

Valle, D., Ziviani, N., and Veloso, A. (2018). Effective

fashion retrieval based on semantic compositional net-

works. In IEEE International Joint Conference on

Neural Networks (IJCNN), pages 1–8.

Wakita, Y., Oku, K., Huang, H.-H., and Kawagoe, K.

(2015). A fashion-brand recommender system using

brand association rules and features. In IEEE Inter-

national Congress on Advanced Applied Informatics,

pages 719–720.

Wang, L. C., Zeng, X. Y., Koehl, L., and Chen, Y. (2015).

Intelligent fashion recommender system: Fuzzy logic

in personalized garment design. IEEE Transactions

on Human-Machine Systems, 45(1):95–109.

Webb, G., Boughton, J., and Wang, Z. (2005). Not so Naive

Bayes: Aggregating one-dependence estimators. Ma-

chine Learning, 58(1):5–24.

Yamaguchi, K., Kiapour, H. H., Ortiz, L. E., and Berg, T. L.

(2012). Parsing clothing in fashion photographs. In

IEEE International Conference on Computer Vision

and Pattern Recognition (CVPR), pages 3570–3577.

Zhao, B., Feng, J., Wu, X., and Yan, S. (2017). Memory-

augmented attribute manipulation networks for inter-

active fashion search. In IEEE International Con-

ference on Computer Vision and Pattern Recognition

(CVPR), pages 6156–6164.

Naive Bayesian Network for Automated, Fashion Personal Stylist

821