Connected Closet

A Semantically Enriched Mobile Recommender System for Smart Closets

Anders Kolstad

Ozlem

Ozg¨obek

, Jon Atle Gulla

and Simon Litlehamar

Department of Computer and Information Science, NTNU, Trondheim, Norway

Accenture AS, Fornebu, Norway

Keywords:

Recommender Systems, Web of Data, Semantic Web, Mobile Systems, Fashion Recommendation, Internet of

Things.

Abstract:

A common problem for many people is deciding on an outﬁt from a vastly overloaded wardrobe. In this

paper, we present Connected Closet, a semantically enriched Internet of Things solution of a smart closet with

a corresponding mobile application for recommending daily outﬁts and suggesting garments for recycling

or donation. This paper describes the whole design and architecture for the system, including the physical

closet, the recommender algorithms, the mobile application, and the backend comprising of microservices

implemented using container technology. We show how users can beneﬁt from the system by supporting them

in organizing their wardrobe, and receiving daily personalized outﬁt suggestions. Moreover, with the system’s

recycling suggestions, the system can be beneﬁcial for the sustainability of the environment and the economy.

1 INTRODUCTION

According to a technical report by the National In-

stitute for Consumer Research, the typical Norwe-

gian has on average 80.4 kg of clothes in their closet

(Klepp and Laitala, 2016). This is equivalent to 359

different clothing items. This suggests that Norwe-

gians have a need for better structuring of the cloth-

ing items in their closets. Moreover, the report states

that 20% of the clothes were never used or only used

a couple of times. This might suggest that the per-

son did not actually like the item they bought from

the clothing retailer. Today, many organizations and

clothing retailers offer checkpoints where people can

delivergarments for recycling or donation. This bene-

ﬁts the sustainability for both the environment and the

economy (Chavan RB, 2014). Moreover, Pruit (2015)

argues that our selection of an outﬁt inﬂuences oth-

ers’ impressions of us and that careful selection of an

outﬁt is of high importance to our social and cultural

lives.

It has long been spoken of the huge amount of data

generated by user-generated content on the Web. Be-

cause of this exponential growth of data, the era of

big data has arised (Jagadish et al., 2014). This huge

growth of data has resulted in information overload.

This implies a clear need for applications able to help

users navigate through the vast amount of content in

a personalized way. Such applications can be made a

reality by recommender systems. Furthermore, this

vast amount of data also calls for a need to struc-

ture the data available online in a meaningful way

(Bizer et al., 2009). As a result, a huge amount of Re-

source Description Framework (RDF) data has been

published as Linked Open Data (LOD) through the

Linking Open Data project

. This kind of data have

huge potential power and recommender systems can

beneﬁt from this data, generating even more accurate

and personalized recommendations (Figueroa et al.,

2015).

In this paper, we present Connected Closet, a se-

mantically enriched Internet of Things (IoT) smart

closet, using clothing items enabled with radio-

frequencyidentiﬁcation (RFID) tags, keeping track of

the clothing items’ usage history. We describe an im-

plementation of a mobile application prototype and

its backend where the users can keep track of what

items currently are in their closets, get recommenda-

tions on what to wear for today, and get recommen-

dations on what to donate and recycle. Furthermore,

we propose three different recommender approaches

in the domain of fashion recommendation.

The work presented in this paper is a collaborative

https://www.w3.org/wiki/SweoIG/TaskForces/

CommunityProjects/LinkingOpenData

298

Kolstad, A., Özgöbek, Ö., Gulla, J. and Litlehamar, S.

Connected Closet - A Semantically Enriched Mobile Recommender System for Smart Closets.

DOI: 10.5220/0006298002980305

In Proceedings of the 13th International Conference on Web Information Systems and Technologies (WEBIST 2017), pages 298-305

ISBN: 978-989-758-246-2

effort between the Smartmedia program

at the Nor-

wegian University of Science and Technology

and

Accenture Norway

. The Smartmedia program fo-

cuses on mobile context-aware recommender systems

research. The goal of this program is to present a

context-aware, personalized news reading experience

based on deep understanding of the textual content of

news articles. Accenture is a leading management,

strategy, consulting, and technology company em-

ploying approximately 384,000 people. Using tech-

nology, Accenture strives to help businesses create

IT applications to deal with rapid changes caused by

an increasingly digital economy. Using their new IT

strategy, they deliver liquid, connected, and intelli-

gent applications to their clients. In this research

project, Accenture demonstrate how to realize the

newIT strategy, and explore applications in their tech-

nology vision by showcasing the advantages of intel-

ligent automation to their clients.

The contributions of this paper are: (1) the archi-

tecture and design of a smart IoT closet; and (2) three

recommendation techniques for recommending items

in the fashion domain.

The rest of the paper is structured as follows. Sec-

tion 2 introduces the background theory. In Section 3,

we give an overview of related work. In Section 4,

we describe the smart closet and the user interface for

the mobile application, followed by a more detailed

description on how the closet is supported by recom-

mendation technology in Section 5. We concludewith

a summary, and discuss future work and possible ben-

eﬁts of the system.

2 BACKGROUND

2.1 Linked Data

Today, most of the data on the Web is uploaded as

HTML documents, or raw dumps such as CSV. This

way of uploading data sacriﬁces much of the Web’s

structure and semantics. Bizer et al. (2009) has

outlined a set of rules for publishing data by using

the Web to create typed links between the data. By

publishing data according to these rules, computer

agents can read and make sense of the data published,

making it easier to gather meaningful information

from the Web. Linked Data published under an open

license is called LOD. An example of LOD is Wiki-

http://research.idi.ntnu.no/SmartMedia/

http://www.ntnu.edu/

https://www.accenture.com/no-en/

data

, which is an open knowledge base that can be

read or edited by any human or computer agent. Us-

ing Wikipedia, Wikivoyage, Wikisource, and others

as its central storage, Wikidata contains millions of

RDF triples, following a Subject, Predicate, Object

structure exposed on its SPARQL endpoint.

2.2 Recommender Systems

The objective of a recommender system is to guide

users in making choices by giving them personal-

ized recommendations of items. Typically, recom-

mender systems are classiﬁed into collaborative ﬁl-

tering and content-based (Jannach et al., 2010). Col-

laborative ﬁltering generate recommendations on the

idea that if some users shared the same interest on

previous items, they will have similar preferences to

other items as well. Content-based recommendations

base its recommendations on item descriptions and a

user proﬁle. A user proﬁle is a set of a user’s prefer-

ences; the recommender system will then recommend

the items that have the most similar item description

to the user proﬁle.

An interesting challenge for all recommender sys-

tems is computing accurate recommendations when

few user ratings are available. This challenge is

known as the sparsity problem (Jannach et al., 2010).

A special case of the sparsity problem is: (a) dealing

with new users who have not yet rated any items; and

(b) how to recommend new items that has not been

rated yet. These two problems are commonly known

as the new-user and new-item cold-start problem.

The majority of today’s recommender systems ad-

dresses recommendations of items in the domains of

movies, books, and music. Different techniques of the

approaches above are well researched and evaluated

in these domains (Bobadilla et al., 2013).

2.3 Internet of Things

IoT is a set of Internet-connected devices embedded

with hardware, software, sensor, actuators, identiﬁers,

and network technologies. These devices collect and

exchange data with each other and other components

on the Internet, generating a vast amount of data ev-

ery day (Gubbi et al., 2013). IoT meets the need for

data-on-demand by intuitive interactions with ubiqui-

tous computing devices. Furthermore, IoT has been

identiﬁed as one of the key trends that organizations

must keep track of to gain competitive advantage, and

that the market adaptation is predicted to take 5–10

years.

https://www.wikidata.org/

Connected Closet - A Semantically Enriched Mobile Recommender System for Smart Closets

299

The applications in the domains that will be and

has been impacted by IoT devices range from control

of home equipment such as refrigerators, to monitor-

ing the water quality in cities.

3 RELATED WORK

Several works have been done on enabling Linked

Data into recommender systems in order to improve

their recommendation algorithms (Figueroa et al.,

2015).

Heitmann and Hayes (2010) describe a recom-

mender system that tries to mitigate i) the new-item

problem; ii) the new user problem; and iii) the spar-

sity problem of recommendations of music by utiliz-

ing Linked Data. They transformed RDF graphs into

a user-item matrix by using data from MySpace and

data about a Wikipedia editor’s homepage. Their re-

sults showed that by enabling Linked Data, the aver-

age precision increased by 14% and the average recall

increased by 33%.

Di Noia et al. (2012) propose a recommender sys-

tem that relies exclusively on information extracted

from the Web of Data. For recommending movies

they propose a content-based recommender system

using the SPARQL endpoints exposed by DBpedia,

LinkedMDB, and Freebase as the base of their rec-

ommender system. To compute similarities between

movies they used the Vector Space Model, represent-

ing the whole RDF as a 3-dimensional matrix where

each slice refers to an ontology property. Given a

property, each movie is seen as a vector. For a given

slice, the similarity is computed between the corre-

lating movie vectors by calculating the cosine angle

between the vectors.

Tomeo et al. (2016) generated a dataset consist-

ing of Facebook likes of music, books, and movies.

First, they mapped the likes to entities in DBpe-

dia to enrich the item proﬁles in the dataset. Then,

they compared the dataset on different graph-based

recommender systems and matrix factorization sys-

tems. Overall, the graph-based algorithm, PathRank

showed the most promising results.

Many prototypes of smart IoT closets for sug-

gesting outﬁts by using RFID technology have been

made in the past (Goh et al., 2011; Ling et al., 2007;

Toney et al., 2006). These prototypes show some very

promising techniques for IoT closets and are built

on the same fundamental techniques as described in

this paper. These techniques involves attaching RFID

tags to the clothing items or hangers, which can be

scanned by a reader in the closet, and then broadcast-

ing a message about the state of the item to a database.

Moreover, similar to our system, some of them also

enable weather data or calendar integration (Schaad

et al., 2016; Liu et al., 2012). As far as we know, none

of them utilize LOD as we do in our system. More-

over, none of them focus on recommending clothing

items that the user might want to donate or recycle.

Research and evaluation of the recommender algo-

rithms in these studies are lacking or has been scoped

out.

Ingvaldsen et al. (2015) propose a recommen-

dation technique for how personalized and location

aware news recommendations can be constructed

based on the users’ contexts. Moreover, they show

how the recommended content can be enriched by us-

ing Wikidata. In our prototype of Connected Closet,

we use similar techniques to combine Wikidata with

context aware user ratings to construct location aware

recommendations based on the weather at the user’s

location.

External

services

Mobile applicationCloset

RFID tag

Cloud

RFID

Figure 1: High level architecture.

4 ARCHITECTURE

The main parts of the prototype are constructed as

follows. As shown in Figure 1, the closet is em-

bedded with a Raspberry PI

, a tiny computer, which

is connected to an RFID reader. When an end-user

touches the RFID tag of a clothing item onto the RFID

reader, the Raspberry PI will broadcast a message

to the backend of the prototype which is constructed

of microservices running in the cloud, each perform-

ing their own designated task. The mobile applica-

tion communicates with the microservices, provid-

ing the end-user with recommendations and inventory

overview. Moreover, the high level architecture con-

tains external services which consists of third party

APIs, such as weather data and LOD.

https://www.raspberrypi.org/

WEBIST 2017 - 13th International Conference on Web Information Systems and Technologies

300

The components of Connected Closet are con-

nected as follows.

4.1 Closet

The computer embedded in the physical closet runs

a Python script listening to scans of clothing items.

When a scanning occurs, the script broadcasts a Mes-

sage Queuing Telemetry Transport (MQTT) message

containing timestamp, item id, user id of the closet

owner, and the status of the clothing item (whether it

is being checked in or out of the closet). Additionally,

by using LED lights and speakers connected to the

computer, the user receives feedback on an item scan.

A red light indicates an insertion, while green light

indicates extraction. This is implemented to main-

tain consistency between the physical clothing items

in the closet and the status of the items stored by

the microservices. To implement this, we used ideas

from Fog Computing (Bonomi et al., 2012) and im-

plemented a local cache database in the embedded

computer that keeps track of the status of the latest

item scans. Additionally, this prevents the Python

script in broadcasting unnecessary messages, such as

double scans.

Figure 2: Prototype of the physical closet.

Figure 2 shows a picture of an early version of

the prototype, including the physical closet embed-

ded with the Raspberry PI. The prototype is built for

demonstration purposes, and to show how human in-

teraction with the closet would work in practice.

4.2 Backend

The backend of the prototype (Figure 3) consists

of ﬁve main components implemented as microser-

vices: the inventory service, the history service,

the catalog service, the recommender service, and

the closet service. All the microservices and their

databases have been implemented using container-

based virtualization with Docker

. This type of

virtualization is an operating-system-level virtualiza-

tion method for running distributed applications with-

out the need for launching an entire virtual ma-

chine. With such container-based microservices,

the whole solution beneﬁts from a horizontal scal-

able architecture composed of small, indepenedent,

and highly coupled components communicating with

each other by means of Representational State Trans-

fer (REST) with the Hypertext Transfer Protocol

(HTTP). The main components of the backend are de-

scribed throughout the section.

History

service

Closet

service

Log

storage

Closet

storage

Catalog

service

Inventory

service

Wikidata

Item

storage

Message

broker

Weather

API

SPARQL

MQTT

Recommender

service

Figure 3: Detailed architecture of the backend.

4.2.1 The Inventory Service

The inventory service is responsible for storing data

about the clothing items registered to an owner of

a closet. The service stores its data in a document-

oriented database. Each owner of a closet is assigned

their own document in this database. The owner’s

clothing items are represented as a list of triples con-

taining the unique id of the clothing item, the article

number of the item, and the status of the item (inside

or outside of closet). Moreover, this service stores

user information such as username and favorite out-

ﬁts. The favorite outﬁts are represented as a list of

tuples containing two items, one top and one bottom.

When the service receives the MQTT message,

the service will then take note of which user the scan-

ning came from and update the user’s document in the

database.

https://www.docker.com/

Connected Closet - A Semantically Enriched Mobile Recommender System for Smart Closets

301

4.2.2 The History Service

The history service is responsible for logging every

scan that occurs in the closet. A time-series database

is connected to this service for storing each log entry.

For each scanning, a record containing the item id, the

timestamp, status of the item, and current temperature

will be saved. Furthermore, this service writes usage

history and temperatures to a database shared with the

recommender and the catalog service.

4.2.3 The Catalog Service

The catalog service is responsible for handling data

about all the different clothing items that are sup-

ported by a Connected Closet. A supported item is an

item connected to an RFID tag with its article number

stored in the database shared with the history and the

recommender service. This service handles match-

ing clothing items and other item properties, such as

color.

Furthermore, this service is set up to communicate

with Wikidata via Wikidata’s SPARQL endpoint, se-

mantically enriching the clothing items in the shared

database. E.g., if one clothing items is registered to

have the color ’navy’ and another item is ’blue’, the

results from Wikidata will include similar description

to both of these colors, making the similarity between

the item descriptions even stronger. This is also done

on other item properties where this is expedient.

4.2.4 The Recommender Service

The recommender service lies in the heart of the

recommendation approaches explained in Section 5.

This service uses the item ratings and the descrip-

tive item data stored in the shared database, called

Item storage. This database is implemented as a graph

database. Figure 4 shows a simpliﬁed example of how

the data is represented in the database.

To realize the recommendation approaches, the

service employs different machine learning libraries.

4.2.5 The Closet Service

The closet service is responsible for providing the

mobile application with meaningful information gath-

ered from the lower levelservices. It generates a set of

the closet overview by joining the data from the cata-

log service and the inventory service. Using weather

data and item status from inventory as input to the

recommender service, the recommender service will

return a list of recommended outﬁts. For getting re-

cycling recommendations from the recommender ser-

vices, it uses data gathered from the history service.

:ClothingItem{

name: ”T-shirt”,

type: ”top”,

color: ”red”,

temp_range: [10, 30]

}

:User{

user_id: 1

}

:ClothingItem{

name: ”Jeans”,

type: ”bottom”,

color: ”blue”,

temp_range: [-5, 20]

}

:USER_RATING{

rating: 0.8

}

:USER_RATING{

rating: 0.4

}

:M A T C H E S _ W I T H {

ma t c h _ w e i g h t : 0 . 4

}

Figure 4: Example of data representation in Item storage.

4.3 Mobile Application

A progressive web application is developed to make

the closet explorable on mobile devices. In this ap-

plication, the user is allowed to view suggestions for

today’s outﬁt, view an inventory of their closet, and

view suggestions on what clothing items to recycle or

donate.

The mobile application communicates with the

closet service by REST through HTTP. A web server

is set up in the middle to handle trafﬁc and connec-

tions using Nginx

4.3.1 User Interface

To view suggestions for today’s outﬁt, the user

chooses the Outﬁt button from the lower menu bar.

Figure 5a shows an illustration of the outﬁt sugges-

tion view. In the top of the view, the weather and

location for the user is displayed. Below is the sug-

gested outﬁt. The user can modify the suggested outﬁt

by clicking on the arrows next to the clothing items of

the outﬁt. If the user wants to go back to what the

system has recommended for today, they can use the

Today’s suggestion button, loading the initial recom-

mendation. The user can save the outﬁt displayed as

a favorite by using the button next to Today’s sugges-

tion. Furthermore, the user can browse through a list

of top-k outﬁt recommendations by swiping up and

down on the screen.

Figure 5b shows an example of a closet overview.

By choosing the My Closet button from the menu, the

overview of the user’s closet is displayed. Here, the

user can browse all the clothing items registered to

their closet and see item status indicated by a closet

icon with a check mark. Moreover,a ﬁltering function

https://nginx.org/en/

WEBIST 2017 - 13th International Conference on Web Information Systems and Technologies

302

(a) Outﬁt suggestion (b) Closet overview (c) Recycling suggestion

Figure 5: Screenshots from the Connected Closet prototype.

is implemented to help the user navigate through the

closet overview more easily.

By choosing the More button from the menu, the

user is displayed with a list of items rarely used and a

suggestion to recycle these items. Figure 5c shows a

suggestion for two items that have not been used for

the past 6 months.

5 FASHION RECOMMENDATION

In this section we describe how the whole system is

supported by recommender technology. We describe

how the item properties needed to do fashion recom-

mendation are determined. Furthermore, we divide

fashion recommendation into outﬁt recommendation

and recycling recommendation.

5.1 Ratings

The user ratings used in fashion recommendation are

determined depending on several context factors. The

context factors for determining the user rating of an

item are:

• Usage history: How often an item has been ex-

tracted from the user’s closet affects the item’s rat-

ing. It is safe to assume that a frequently used item

is an item that the user likes. Therefore, an item

that is used on a weekly basis will have a high

rating.

• Current season: Some clothing items are seasonal.

E.g., a winter jacket will have higher rating during

the winter, and a low rating in the summer.

• Weather: Much like season, some items are

weather dependent, e.g., rain coat. The rating of

such items will therefore be affected by the daily

weather.

• Taste proﬁle: As the user saves outﬁts as favorites,

the items in the outﬁts will increase their user rat-

ings.

Using these factors as input, the rating of an item is

set using a 10-star rating scale.

5.2 Matches and Suitable Temperatures

For determining matching clothing tops and bottoms

in the catalog service, we calculate a weight between

the two items and assign it to their edge. Initially, all

tops and bottoms match each other with a weight of

0.0. When two items are checked out of the closet

during the same 2-hour time period, the weight be-

tween these items increase with α. Furthermore,

when a user saves an outﬁt as a favorite using the

mobile application, the weight increases by β. The

matching weight cannot exceed 1.0.

All clothing items are saved in the catalog service

with a suitable temperature range property. The suit-

Connected Closet - A Semantically Enriched Mobile Recommender System for Smart Closets

303

able temperature range is the range of temperatures in

which a clothing item is comfortable to wear. This

range is determined by the average temperature of all

the checkouts of a clothing item, calculated by For-

mula 1:

st(i) =

(

∑

j=1

i, j,temp

) ± δ

◦

C, (1)

where N is the number of checkouts of clothing item i

in C

and δ is a constant determining the length of the

range.

5.3 Outﬁt Recommendation

In our system, two approaches for outﬁt recommen-

dation are implemented. The ﬁrst approach uses

ideas from collaborative ﬁltering, while the second

approach is a pure content-based approach enabled

with LOD.

For an item to be included in a recommended out-

ﬁt it must be: (1) inside the closet; and (2) the current

temperature must be inside the items suitable temper-

ature range.

5.3.1 Outﬁt-item Matrix

In the ﬁrst approach, we transform the outﬁts saved

as favorites by the end-users into an outﬁt-item ma-

trix. In this matrix, each column represents a favorite

outﬁt of an end-user. The rows represent every item

supported by Connected Closet and that is part of a

user’s favorite outﬁt. Table 1 shows an example of an

outﬁt-item matrix with three outﬁts and four items.

All outﬁts is associated with a weight, e.g, w

. These

weights are based on number of likes of the outﬁt and

are used to determine the strength of the outﬁt, mak-

ing it easier to neglect outﬁts favorited by few users.

Using this matrix as training data, different classiﬁca-

Table 1: Example of an outﬁt-item matrix.

Outﬁt 1 Outﬁt 2 Outﬁt 3

Item 1 × ×

Item 2 ×

Item 3 × ×

Item 4 ×

tion algorithms are applied to classify outﬁts as good

or neutral. A good outﬁt means an outﬁt that can be

recommended to the user. While outﬁts classiﬁed as

neutral are outﬁts that the users either does not like or

has not been rated yet, and will therefore not be rec-

ommended to the users. In our method, we ﬁrst create

outﬁts combinations of the items that ﬁt our inclusion

criteria. Then, we input these outﬁts into the classiﬁ-

cation model. Outﬁts that are classiﬁed as good will

then be recommended to the user.

5.3.2 Vector Space Model

For our content-based approach, we use a vector

space model similar to the one proposed by Di Noia

et al. (2012). Using the user ratings stored in the item

storage, we build a user proﬁle consisting of clothing

items with a rating above λ, using Formula 2:

profile(u) = {c

| r

> λ}, (2)

where r

is rating of clothing item c

for user u.

We then generate a ranked list of all the clothing

items in the user’s closet using Formula 3:

¯r(u, c

) =

∑

∈profile(u)

sim(c

, c

)

|profile(u)|

, (3)

where sim(c

, c

) is a similarity measure between the

vectors representing the clothing items c

and c

We then ﬁlter out a list of top-k outﬁts based on

our inclusion criteria and the match weight between

the tops and bottoms.

5.4 Recycling Recommendation

For recommending items that may be of interest for

recycling by the end-user, the system returns a list of

the three lowest rated items that have been rarely used

in the past 12 months. By using a time period of min-

imum 12 months it safe to recommend items that are

also seasonal.

An identiﬁed problem with the proposed tech-

nique is the new user cold-start problem. The owner

of a Connected Closet should be able to get relevant

recycling recommendations from the day that they

acquire the closet. This problem, and other tech-

niques for recycling recommendationwill be explored

in later research.

6 CONCLUSION AND FUTURE

WORK

The described prototype is an ongoing project with

some development still remaining. A full evaluation

and validation will be performed in later research. In

this paper we have proposed a novel IoT system for

doing fashion recommendation using modern tech-

nologies, such as LOD, microservices, containers,

and progressive web apps. The proposed system can

WEBIST 2017 - 13th International Conference on Web Information Systems and Technologies

304

guide users in making daily outﬁt selections and ef-

ﬁciently organize their wardrobe in an environment-

friendly way. Furthermore, if the system were to be

integrated in a clothing retailer’s supply chain and

used by their customers, the retailer could generate

targeted ads and provide relevant recommendations to

their customers.

As an initial evaluation, the prototype has been

showcased at various IT conferences and events.

At these events, the participants have been given a

demonstration of the prototype and have had the op-

portunity to try out the prototype for themselves. The

response from the participants has been positive, and

many participants have expressed that they would

beneﬁt from such a system in their everyday lives.

Future work will be devoted to gathering data for a

dataset that can be used for accuracy evaluation of the

recommendation approaches. Moreover, we aim to do

an user-centric evaluation of the recommender system

in order to evaluate the user satisfaction. Some other

topics we would like to research further and include

in our system are occasion-based outﬁt recommenda-

tions and recommendations from retailers.

ACKNOWLEDGEMENTS

This work is composed by a research cooperation that

was established subsequently of a summer internship

at Accenture, where the idea and the ﬁrst prototype

of Connected Closet originally was developed. The

authors would like to thank everyone that was in-

volved in the internship for their contributions prior

this work.

REFERENCES

Bizer, C., Heath, T., and Berners-Lee, T. (2009). Linked

data-the story so far. Semantic Services, Interoperabil-

ity and Web Applications: Emerging Concepts, pages

205–227.

Bobadilla, J., Ortega, F., Hernando, A., and Gutirrez, A.

(2013). Recommender systems survey. Knowledge-

Based Systems, 46:109 – 132.

Bonomi, F., Milito, R., Zhu, J., and Addepalli, S. (2012).

Fog Computing and Its Role in the Internet of Things.

In Proceedings of the First Edition of the MCC Work-

shop on Mobile Cloud Computing, MCC ’12, pages

13–16, New York, NY, USA. ACM.

Chavan RB (2014). Environmental Sustainability through

Textile Recycling. Journal of Textile Science & Engi-

neering, pages 1–6.

Di Noia, T., Mirizzi, R., Ostuni, V. C., Romito, D., and

Zanker, M. (2012). Linked Open Data to Support

Content-based Recommender Systems. In Proceed-

ings of the 8th International Conference on Semantic

Systems, I-SEMANTICS ’12, pages 1–8, New York,

NY, USA. ACM.

Figueroa, C., Vagliano, I., Rocha, O. R., and Morisio,

M. (2015). A systematic literature review of Linked

Data-based recommender systems. Concurrency and

Computation: Practice and Experience, 27(17):4659–

4684.

Goh, K. N., Chen, Y. Y., and Lin, E. S. (2011). Devel-

oping a smart wardrobe system. In 2011 IEEE Con-

sumer Communications and Networking Conference

(CCNC), pages 303–307.

Gubbi, J., Buyya, R., Marusic, S., and Palaniswami, M.

(2013). Internet of Things (IoT): A vision, architec-

tural elements, and future directions. Future Genera-

tion Computer Systems, 29(7):1645–1660.

Heitmann, B. and Hayes, C. (2010). Using Linked Data to

Build Open, Collaborative Recommender Systems. In

AAAI spring symposium: linked data meets artiﬁcial

intelligence, pages 76–81.

Ingvaldsen, J. E., Ozgobek, O., and Gulla, J. A. (2015).

Context-aware user-driven news recommendation. In

3rd International Workshop on News Recommenda-

tion and Analytics (INRA 2015). Ceur-ws Vol-1542.

Jagadish, H. V., Gehrke, J., Labrinidis, A., Papakonstanti-

nou, Y., Patel, J. M., Ramakrishnan, R., and Shahabi,

C. (2014). Big Data and Its Technical Challenges.

Commun. ACM, 57(7):86–94.

Jannach, D., Zanker, M., Felfernig, A., and Friedrich, G.

(2010). Recommender Systems: An Introduction.

Cambridge University Press.

Klepp, I. G. and Laitala, K. (2016). Clothing consumption

in Norway. Technical report 2, Oslo and Akershus

University College, Oslo.

Ling, S., Indrawan, M., and Loke, S. W. (2007).

RFID-Based User Proﬁling of Fashion Preferences:

Blueprint for a Smart Wardrobe. Int. J. Internet Pro-

toc. Technol., 2(3/4):153–164.

Liu, S., Feng, J., Song, Z., Zhang, T., Lu, H., Xu, C., and

Yan, S. (2012). Hi, Magic Closet, Tell Me What to

Wear! In Proceedings of the 20th ACM International

Conference on Multimedia, MM ’12, pages 619–628,

New York, NY, USA. ACM.

Pruit, J. C. (2015). Getting Dressed. Popular Culture as

Everyday Life, page 155.

Schaad, P., Basler, S., Medini, M., Wissler, I., Brkle, T.,

and Lehmann, M. (2016). The Intelligent Wardrobe

. Studies in Health Technology and Informatics,

225:213–217.

Tomeo, P., Fernndez-Tobas, I., Di Noia, T., and Cantador, I.

(2016). Exploiting Linked Open Data in Cold-start

Recommendations with Positive-only Feedback. In

Proceedings of the 4th Spanish Conference on Infor-

mation Retrieval, CERI ’16, pages 11:1–11:8, New

York, NY, USA. ACM.

Toney, A. P., Thomas, B. H., and Marais, W. (2006). Man-

aging Smart Garments. In 2006 10th IEEE Interna-

tional Symposium on Wearable Computers, pages 91–

94.

Connected Closet - A Semantically Enriched Mobile Recommender System for Smart Closets

305