An Agent based Platform for Resources Recommendation in

Internet of Things

Agostino Forestiero

and Giuseppe Papuzzo

Institute for High Performance Computing and Networking, National Research Council of Italy,

Via P. Bucci 8-9C, Rende (CS), Italy

Keywords:

Multiagent Systems, Internet of Things, Word Embedding, Recommendation Systems.

Abstract:

Internet of Things (IoT) paradigm aims to bridge the gap between physical and cyber world allowing a deeper

understanding of users in terms of preferences and behaviors. User devices and services interact and maintain

relations which need of effective and efﬁcient selection/suggestion approaches to better meet users’ interests.

Recommendation systems provide a set of signiﬁcant and useful suggestions for users and systems with given

characteristics. This paper introduces the design of an agent based platform for building a distributed rec-

ommendation system in IoT environment. Internet of Things objects (devices, sensors, services, etc.) are

represented with vectors obtained through the Doc2Vec model, a neural word embedding approach able to

capture the semantic context representing documents in dense vectors. The vectors are managed by cyber

agents that, performing simple and local operations, organize themselves exploiting the vector values. The

outcome is the emergence of an organized overlay-network of cyber agents that allows to obtain an efﬁcient

recommender system of IoT services. Preliminaries results conﬁrm the validity of the approach.

1 INTRODUCTION

Several objects interconnected with each other

achieving a common purpose represent the Internet

of Things (IoT) paradigm (Atzori et al., 2010). The

number of objects and the amount of data gener-

ated by IoT infrastructures are growing hugely and

this makes the traditional manage mechanisms in-

adequate. Intelligent and automated approaches are

needed to support decision makers due to the dynamic

nature of smart objects, devices and services, involved

in the IoT systems. Systems able to perform “things

recommendation” is a crucial step to take full advan-

tage of the IoT. Recommendation systems are an im-

portant research topic and several works have been

proposed both in the industry and academia. These

systems allow to individuate a list of useful items for

the users in a given context. The usefulness of an item

or product or service is generally represented by a

“rating”, which indicates how much a given user likes

a particular item. The items with a high rating value

are presented as recommendations for the user.

Recommendation systems can be categorized as

(Balabanovi´c and Shoham, 1997): (i) Collaborative

https://orcid.org/0000-0002-3025-7689

https://orcid.org/0000-0003-3961-3966

Filtering (CF), an item is recommended to the user ac-

cording to the past ratings of all users. The approach

evaluates the utility of the item i for the user u by es-

timating the usefulness assigned to item i by the users

v who are “similar” to user u; (ii) Content-based rec-

ommending an item is recommended if it is similar to

items that the user has chosen in the past. Information

retrieval (IR) techniques address this problem, where

the content associated can be handled as a query, and

the unrated documents marked with a similarity value

to this query. Otherwise, the documents can be con-

verted into word vectors, and then averaged to obtain

a prototype vector of each category for a user, as re-

ported in (Lang, 1995); and (iii) Hybrid approaches

in which collaborative and content-based approaches

are combined. The similarity between two users can

be computed using various approaches, but the most

popular are correlation (Resnick et al., 1994) and co-

sine similarity (Breese et al., 1998). Collaborative and

content-based approaches use the same cosine mea-

sure from information retrieval. But, in content-based

recommendation systems measures the similarity be-

tween vectors of weights, whereas, in collaborative

systems measures the similarity between vectors of

the actual ratings speciﬁed of the users. Other ap-

proaches to the recommendation consist in handling

Forestiero, A. and Papuzzo, G.

An Agent based Platform for Resources Recommendation in Internet of Things.

DOI: 10.5220/0009573207750779

In Proceedings of the 22nd International Conference on Enter prise Information Systems (ICEIS 2020) - Volume 1, pages 775-779

ISBN: 978-989-758-423-7

775

of the problem as a classiﬁcation task. Each pattern

represents the content of an item, and a user’s past

ratings are used as labels for these patterns. For ex-

ample, text from ﬁelds such as title, author, synopses,

reviews, and subject terms are used by (Mooney and

Roy, 2000) to recommend books.

Several classiﬁcation algorithms have been also

used to content-based recommend: decision trees, k-

nearest neighbor, and neural networks (Pazzani and

Billsus, 1997). The heterogeneity of possible scenar-

ios, arising from the massive deployment of an enor-

mous amount of smart objects, imposes the use of so-

phisticated and innovativemodels and algorithms. We

propose an advanced version, enriched with seman-

tic properties, of the agent-based algorithm for build-

ing a things recommendation system introduced in

(Forestiero, 2017), in which bio-inspired agents work

together in order to obtain a common purpose that is

the organization of distributed resources. Agents are:

extensible, they can be created or modiﬁed; stable,

when an agent is out of services, other agents share

tasks and ensure the continuity of services; and inde-

pendent, they are running without user intervention.

With characteristics as extensibility, stability and au-

tonomy, they can be more adequate in a dynamic sys-

tem because they automatically adapt to environments

change (Selmi et al., 2014).

The algorithm proposed is able to organize things

of an IoT environment in order to improve recom-

mendation operations. In particular, each smart ob-

ject is associated with a “cyber agent”, which repre-

sents it in a cyber layer. The cyber layer is a vir-

tual layer in which the cyber agents can“collaborate”

among them, in a peer to peer fashion (Forestiero

et al., 2008a), in order to obtain a common goal and

improving the performances of the system (Forestiero

et al., 2008b)(Forestiero et al., 2005). Vectors of real

numbers, are exploited to describe IoT objects. The

vector can have different meanings, for example: the

presence or absence of a given characteristic or it can

be the result of a hash function locality preserving

so that similar vectors are assigned to things with

similar characteristics. In peer to peer systems, in-

deed, metadata representing the content are often in-

dexed through bit vectors, or keys, which can have

different meanings. One is that each bit represents

the presence or absence of a given topic (Crespo and

Garcia-Molina, 2002) (Platzer and Dustdar, 2005):

this method is particularly adapt for contents like doc-

uments, because it is possible to identify the differ-

ent topics existing in the documents. Alternatively, a

metadata can be mapped through a hash function into

a binary vector. The hash function haveto be designed

locality preserving (Cai et al., 2004) (Oppenheimer

et al., 2005), thus, neighbor vectors are assigned to

contents with neighbor/similar characteristics. Simi-

larity measure can be the cosine of the angle or the

Euclidean distance between the bit vectors.

In our approach, the Doc2Vec model (Le and

Mikolov, 2014), able to represent documents in dense

vectors, also capturing the semantic, is exploited to

map smart objects. The cyber agents organize them-

selves based on the similarity of its vector with the

wished IoT device/service. The outcome of the algo-

rithm is a logically sorted list of cyber agents based

on the similarity with the target IoT device/service,

where the distance from the target IoT device/service

increase with the distance from the initial position of

the list. Thanks to this organized list, the sugges-

tion operations become faster, because we can ﬁnd

similar, probably useful and recommendable vectors

(smart objects) in the ﬁrst positions of the list.

2 SMART RECOMMENDATION

ARCHITECTURE

The aim is to design a platform able to provide useful

resource suggestions in Internet of Things. To achieve

this objective, a semantic multiagent algorithm was

designed and implemented. Physical devices, sensors,

services, etc. are represented by vectors obtained

through the Doc2Vec neural model (Le and Mikolov,

2014) applied to the metadata (text) describing them.

Doc2Vec is an unsupervised algorithm to generate

vectors starting from sentences/documents based on

Word2Vec, a word embedding approach which can

generate vectors starting from words. Word2Vec (Le

and Mikolov, 2014) is a two layer artiﬁcial neu-

ral network used to process text to learn relation-

ships between words within a text corpus. Word2Vec

takes as its input a large corpus of text and pro-

duces a high-dimensional space (typically of several

hundred dimensions), with each unique word in the

corpus being assigned a corresponding vector in the

space. This “word embedding” approach is able to

capture multiple different degrees of similarity be-

tween words. To create the model of relationships

between the words, a particular grouping of text or

documents is fed to the Word2Vec process, which is

called the training corpus. Word2Vec builds a vocab-

ulary exploiting a corpus and, by training a neural

network with three levels, learns the word represen-

tations. Word2Vec proposes two kind of models: (i)

Continuous Bag of Words (CBOW) that learns the

representations by predicting the target word based on

its context words; and (ii) Skip-gram, that learns rep-

resentations by predicting each of the context words

ICEIS 2020 - 22nd International Conference on Enterprise Information Systems

776

Figure 1: Logical layers of the infrastructure. The Overlay

Layer depicts the outcome of the algorithm.

based on the target word. So, one has to choose one

of the architectures and set values for hyper parame-

ters like embedding size, context size, minimum fre-

quency for a word to be included in the word vocab-

ulary to generate the word embeddings from a large

corpus of unlabeled data. The distance/similarity

between two IoT devices/services can be computed

through the cosine distance/similarity between the

vectors representing the description. Given two IoT

devices/services, the cosine measure utilized to com-

pute the similarity between them is reported in for-

mula (1). The cosine-based method uses two vectors

in n-dimensional space to represent the users u and v,

and n will be |I|. The cosine of the angle between two

vectors, as reported in formula (1), can be computed

to measure the similarity between them, where~u·~v is

the dot-product between the vectors~u and~v.

Similarity(~u,~v) = cos(~u,~v) =

~u·~v

|~u| × |~v|

∑

i=1

∑

i=1

∑

i=1

(1)

Vectors obtained with Doc2Vec library and describ-

ing IoT devices/services, are assigned to cyber agents

that, in fully distributed and self-organizing modal-

ity, exploiting only local information, organize them-

selves in order to improve recommendation opera-

tions. The outcome of the algorithm is a sorted over-

lay network of cyber agents, organized on the basis of

the similarity with the wished IoT device/service.

Figure 1 reports a logical architecture of the algo-

rithm. The infrastructure is organized in three main

layers: (i) a physical layer, composed of a collec-

tion of IoT devices/services connected among them

through local area networks and able to collect data

and information coming from the physical world; (ii)

a cyber layer, consisting of a group of cyber agents

representing the IoT devices/services. In this layer,

the cyber agents work and collaborate together fol-

lowing the steps of the algorithm in order to pro-

duce an overlay network able to provide useful and

relevant information for recommendation operations;

and (iii) an overlay layer, in which an overlay net-

work of cyber agents emerges as outcome of the al-

gorithm. The topology of the emerging overlay de-

pends on the strategy of the algorithm performed by

the cyber agents. In this ﬁgure, a similarity ordered

list of cyber agents emerges. This list allows to select

a given number of cyber agents, representing IoT de-

vices/services, that can be suggested to the user and

probably useful and relevant because they are very

similar to the wished IoT device/service.

2.1 Agent based Algorithm

In the infrastructure proposed, each IoT de-

vice/service is associated with a cyber agent that rep-

resents it in the cyber layer. The cyber agents per-

form autonomously, in a self organizing fashion and

exploiting only local information, all steps of the al-

gorithm in order to obtain an overlay network use-

ful for recommendation operations. The topology

of the emerging overlay depends of the algorithm

strategy. In fact, on the basis of the selection pol-

icy of the neighbors of each cyber agent, it will

emerge a given topology. In this case, the cyber

agents execute a set of steps in order to obtain an

ordered list on the basis of the similarity value with

a given IoT device/service. The ﬁrst element of the

list will be the IoT device/service, among all, hav-

ing the highest value of similarity with the wished

IoT device/service, the second element is the cyber

agent having the second highest value of the simi-

larity, and so on. The steps performed by each cy-

ber agent A

cyber

having vector V

in order to achieve

the algorithm are reported in Algorithm 1. Here,

agents

list containing the linked cyber agents with

vector value higher than Vc and L

agents

list contain-

ing the linked cyber agents with vector value lower

than Vc. The lists are computed through the func-

tion computeList(H

agents

. The function identi-

fyMaxSubmax(H

agents

) returns the linked cyber agents

with the vector value having the maximum and sub-

maximum similarity value with the current cyber

agent, while the function identifyMinSubmin(L

agents

provides the linked cyber agents with the vector value

having the minimum and sub-minimum similarity

value with the current cyber agent. The function cre-

ateLink(ca

,ca

) generates a virtual link between cy-

ber agent ca

and cyber agent ca

. In detail, to ca

notiﬁed to add ca

in its neighbors list and to ca

notiﬁed to add ca

in its neighbors list. The removal

of the cyber agent ca as neighbor of the current cyber

agent is performed by means of function remove(list,

An Agent based Platform for Resources Recommendation in Internet of Things

777

Algorithm 1: Steps of the algorithm.

while true do

computeList(H

agents

);

if H

agents

.length() > 1 then

[max,submax]=identifyMaxSubmax(H

agents

);

createLink(max, submax);

remove(H

agents

, max);

end

else if L

agents

.length() > 1 then

[min,submin]=identifyMinSubmin(L

agents

);

createLink(min, submin);

remove(L

agents

, min);

end

ca) which simply delete ca from the neighbors list.

When this starting phase ﬁnishes, the last cy-

ber agent contained in the list L

agents

, represents

the linked cyber agent with the highest vector value

among all linked cyber agents with the vector value

lower than V

; while, the H

agents

list contains the

linked cyber agent with the lowest vector value among

all linked cyber agents with vector the value higher

than A

cyber

. At a steady situation i.e. after a transi-

tion phase, each cyber agent is connected, with virtual

or real link, with two cyber agents: the cyber agent

having the vector value immediately less similar and

the cyber agent with the vector value of immediately

more similar of the all cyber agents. It is possible to

design a smart recommendation mechanism thanks to

the ordering achieved by the algorithm. In fact, the

cyber agents organize themselves based on the sim-

ilarity with the wished IoT device/service and, once

the algorithm ends, we can select a given number of

IoT devices/service starting from the head of the list.

Thanks to the organization, the suggestion provided

are very similar and, more probably, useful for the

user.

3 EXPERIMENTAL RESULTS

In order to investigate its effectiveness, a Java sim-

ulator was implemented in which the characteristics

of real scenario are careful considered. The simula-

tor allows to build a set of IoT devices/services ran-

domly connected to a given mean number of others

IoT devices/services (neighbors), in order to create a

group of objects connected through a local area net-

work. Each IoT device/services is described thorough

a metadata: a text ﬁle containing a detailed descrip-

tion of the characteristics of the objects. By exploit-

ing the Doc2Vec library, a cyber agent, responsible

of an IoT service/device, can obtain a vector repre-

senting it. The cyber agents following autonomously

Figure 2: The average number of messages managed by

each cyber agent to obtain the logical sorting, for different

value of mean number of linked neighbors.

Figure 3: The total number of messages exchanged by all

cyber agents per step to obtain the sorting. The number of

involved cyber agents is set to 5000.

the steps of the algorithm organize in order to obtain

the overlay useful to improve recommendation oper-

ations. Figure 2 shows the mean number of message

managed by each cyber agent, for different number

of IoT devices/services. The experiments were exe-

cuted for different values of MeanNgh, that is the av-

erage number of connections/neighbors of every cy-

ber agent.

It can be noticed how with a limited number of

messages, the algorithm, reaches a stable situation

and an useful overlay is obtained. In Figure 3 the to-

tal number of messages exchanged by all cyber agents

per step to achieve the global organization, when the

number of the cyber agents involved in the process in

ﬁxed to 5000, is shown.

Notice that the number of messages decreases ex-

ponentially and the algorithm converges in a ﬁnite

number of steps. Successively the worst case to ob-

tain the organization, i.e. the maximum possible num-

ber of steps needed to each cyber agent to individu-

ICEIS 2020 - 22nd International Conference on Enterprise Information Systems

778

Figure 4: The maximum value of possible steps (worst case)

necessary to obtain the logical sorting.

ate its neighbors, is shown in Figure 4. Even in this

case, the experiments were executed for different val-

ues of the average number of connections/neighbors

of cyber agent, when the number of the involved cy-

ber agents changes. The interesting result is that the

maximum number of steps needed to each cyber agent

to individuate its neighbors is always very low and,

anyway, limited to ﬁrst application of the algorithm

in which no one previous sorting exists.

4 CONCLUSIONS

The design of a multiagent platform for building a

recommendation system in Internet of Things boast-

ing semantic features, was presented. This platform

relies on an agent based algorithm allowing discov-

ery and recommendation operations faster. IoT ob-

jects are represented through vectors obtained by ex-

ploiting a word embedding library able to capture

the semantic characteristics. Intelligent agents au-

tonomously execute a distributed algorithm allowing

to bring out an useful overlay for recommendationop-

erations. Experimental results show as the algorithm,

pillar of the platform proposed, enables an effective

reorganization of IoT services/devices, allowing very

encouraging performance for recommendation opera-

tions.

REFERENCES

Atzori, L., Iera, A., and Morabito, G. (2010). The internet of

things: A survey. Computer networks, 54(15):2787–

2805.

Balabanovi´c, M. and Shoham, Y. (1997). Fab: content-

based, collaborative recommendation. Communica-

tions of the ACM, 40(3):66–72.

Breese, J. S., Heckerman, D., and Kadie, C. (1998). Empir-

ical analysis of predictive algorithms for collaborative

ﬁltering. In Proceedings of the Fourteenth conference

on Uncertainty in artiﬁcial intelligence, pages 43–52.

Morgan Kaufmann Publishers Inc.

Cai, M., Frank, M., Chen, J., and Szekely, P. (2004). Maan:

A multi-attribute addressable network for grid infor-

mation services. Journal of Grid Computing, 2(1):3–

14.

Crespo, A. and Garcia-Molina, H. (2002). Routing indices

for peer-to-peer systems. In Distributed Computing

Systems, 2002. Proceedings. 22nd International Con-

ference on, pages 23–32. IEEE.

Forestiero, A. (2017). Multi-agent recommendation sys-

tem in internet of things. In Proceedings of the

17th IEEE/ACM International Symposium on Cluster,

Cloud and Grid Computing, CCGrid ’17, pages 772–

775, Piscataway, NJ, USA. IEEE Press.

Forestiero, A., Mastroianni, C., and Spezzano, G. (2005). A

multi agent approach for the construction of a peer-to-

peer information system in grids. Self-Organization

and Autonomic Informatics (I), 135:220–236.

Forestiero, A., Mastroianni, C., and Spezzano, G. (2008a).

Building a peer-to-peer information system in grids

via self-organizing agents. Journal of Grid Comput-

ing, 6(2):125–140.

Forestiero, A., Mastroianni, C., and Spezzano, G. (2008b).

Reorganization and discovery of grid information with

epidemic tuning. Future Generation Computer Sys-

tems, 24(8):788–797.

Lang, K. (1995). Newsweeder: Learning to ﬁlter netnews.

In In Proceedings of the Twelfth International Confer-

ence on Machine Learning, pages 331–339. Citeseer.

Le, Q. and Mikolov, T. (2014). Distributed representations

of sentences and documents. In International Confer-

ence on Machine Learning, pages 1188–1196.

Mooney, R. J. and Roy, L. (2000). Content-based book rec-

ommending using learning for text categorization. In

Proceedings of the ﬁfth ACM conference on Digital li-

braries, pages 195–204. ACM.

Oppenheimer, D., Albrecht, J., Patterson, D., and Vahdat,

A. (2005). Design and implementation tradeoffs for

wide-area resource discovery. In Proc. of the 14th

IEEE International Symposium on High Performance

Distributed Computing HPDC 2005, Research Trian-

gle Park, NC, USA.

Pazzani, M. and Billsus, D. (1997). Learning and revis-

ing user proﬁles: The identiﬁcation of interesting web

sites. Machine learning, 27(3):313–331.

Platzer, C. and Dustdar, S. (2005). A vector space search en-

gine for web services. In Web Services, 2005. ECOWS

2005. Third IEEE European Conference on, pages 9–

pp. IEEE.

Resnick, P., Iacovou, N., Suchak, M., Bergstrom, P., and

Riedl, J. (1994). Grouplens: an open architecture for

collaborative ﬁltering of netnews. In Proceedings of

the 1994 ACM conference on Computer supported co-

operative work, pages 175–186. ACM.

Selmi, A., Brahmi, Z., and Gammoudi, M. (2014). Multi-

agent recommender system: State of the art. In In

Proc. of the 16th International Conference on Infor-

mation and Communications Security.

An Agent based Platform for Resources Recommendation in Internet of Things

779