Attentional Neural Mechanisms for Social Recommendations in

Educational Platforms

Italo Zoppis

1 a

, Sara Manzoni

1 b

, Giancarlo Mauri

1 c

, Ricardo Anibal Matamoros Aragon

Luca Marconi

and Francesco Epifania

Department of Computer Science, University of Milano Bicocca, Milano, Italy

Social Things srl., Milano, Italy

Keywords:

Social Networks, WhoTeach, Social Recommendations, Graph Attention Networks.

Abstract:

Recent studies in the context of machine learning have shown the effectiveness of deep attentional mecha-

nisms for identifying important communities and relationships within a given input network. These studies

can be effectively applied in those contexts where capturing speciﬁc dependencies, while downloading useless

content, is essential to take decisions and provide accurate inference. This is the case, for example, of current

recommender systems that exploit social information as a clever source of recommendations and / or expla-

nations. In this paper we extend the social engine of our educational platform “WhoTeach” to leverage social

information for educational services. In particular, we report our work in progress for providing “WhoTeach”

with an attentional-based recommander system oriented to the design of programmes and courses for new

teachers.

1 INTRODUCTION

Modern Recommender Systems (RS) use social in-

formation to identify the interest of a target user

and provide reliable suggestions (Bonhard and Sasse,

2006). These “social” recommendations extend tradi-

tional approaches (e.g., collaborative ﬁlter) by obtain-

ing compelling information from complex networks

of users and items (Zhou et al., 2012) to identify per-

sonal interests and implicitly forecast preferences on

available items (Gupta et al., 2013; Schafer et al.,

2007).

In this context, explainable AI is aimed at pro-

viding intuitive explanations for the suggestions and

recommendations given by the algorithm (Zhang and

Chen, 2018). Basically they try to address the prob-

lem of why certain recommendations are suggested

by the applied models. The results of this research

are progressively shortening the distance between so-

cial networks and recommender systems, changing

the way people interact and the content they can share

(Zhou et al., 2012). At the same time, different at-

tempts in current deep learning literature try to extend

https://orcid.org/0000-0001-7312-7123

https://orcid.org/0000-0002-6406-536X

https://orcid.org/0000-0003-3520-4022

deep techniques to deal with social data, recommen-

dations and explanations. Initial work in this context

used recursive networks to process structured data

such as direct acyclic graphs (Frasconi et al., 1998;

Sperduti and Starita, 1997). More recently, Graph

Neural Networks (GNNs) and others machine learn-

ing techniques (Dondi et al., 2019; Dondi et al., 2016;

Zoppis et al., 2019b) have been introduced as a gen-

eralization of recursive networks capable of handling

more general classes of graphs (Gori et al., 2005;

Scarselli et al., 2008). In this regard, emerging re-

searches on deep architectures focus on how to bring

out relevant parts of a network to perform a given task

(Veli

ckovi

c et al., 2017). Technically, this approach

is known as “attentional mechanism” for graphs, or

“Graph attention networks” (GATs). Introduced for

the ﬁrst time in the deep learning community in or-

der to access important parts of the data (Bahdanau

et al., 2014), the attention mechanism has recently

been successful for the resolution of a series of ob-

jectives (Lee et al., 2018). For example, it is worth

to cite (Chen et al., 2018) where explainable sequen-

tial recommendations are extracted due to memory

networks. Another interesting approach comes from

capsule networks (Li et al., 2019), namely neural net-

works empowered with capsule structures to manage

hierarchies.

Zoppis, I., Manzoni, S., Mauri, G., Aragon, R., Marconi, L. and Epifania, F.

Attentional Neural Mechanisms for Social Recommendations in Educational Platforms.

DOI: 10.5220/0009568901110117

In Proceedings of the 12th International Conference on Computer Supported Education (CSEDU 2020) - Volume 1, pages 111-117

ISBN: 978-989-758-417-6

111

In this article, focusing on these researches, we

extend the social engine of our educational platform

(Zoppis et al., 2019a) called “WhoTeach” with a

graph attentional mechanism aiming to provide so-

cial recommendations for the design of new didacti-

cal programmes and courses. More in details, we de-

scribe the “WhoTeach” platform and its social engine

in Sec. 2. In Sec. 3 we consider the main theoreti-

cal aspects used in our framework. We introduce the

attentional based architecture applied in this work in

Sec. 4. In Sec. 5 we report the numerical experiments

on a public dataset. Finally, we conclude the paper in

Sec.6 by describing future directions of this research.

2 WhoTeach

WhoTeach is an innovative e-learning system, aimed

at promoting the development of customized learn-

ing and training paths by aggregating and dissem-

inating knowledge created and updated by experts.

WhoTeach is conceived as a Social Intelligent Learn-

ing Management System (SILMS) and it is structured

around three main components:

1. The Recommender System, which helps experts

and teachers to quickly assemble high-quality

contents into courses: thanks to an intelligent

analysis of available material, it suggests teach-

ers the best resources, in any format, to include.

Moreover, it helps students, employees and work-

ers to acquire intelligent suggestions. This way,

they can ﬁnd customized courses in real time, tak-

ing into consideration their personal proﬁle, ob-

jectives and ambitions, along with rules and crite-

ria deﬁned by their belonging organization.

2. The ”Knowledge co-creation Social Platform”,

which is a technological infrastructure based on

an integrated and highly interactive social net-

work. This component allows learners to interact

and cooperate in the learning path, exchanging in-

formation and mutual advice about the contents.

That ampliﬁes their learning effort and motivates

them to get the best from it. Moreover, it provides

experts and teachers with feedback from learners

about clarity and comprehension of content and

with high-quality comments from domain experts

and other teachers (like a peer review). Conse-

quently, it enables the creation and the exchange

of knowledge inside organizations at no additional

cost. All is done in an informal, unstructured, in-

teractive and pervasive way.

3. The content’s repository where to upload con-

tents from any course or training material, either

proprietary or open, like the ones promoted by

the European Union (for example EMMA or the

Open Education portal). Thanks to the use of

metadata, the system can build a new, original

and high-quality course because WT can identify

the right information needed and obtain it from

any course (internal e/o external) already avail-

able. This mechanism enriches the contents’ of-

fer without additional cost, increasing the efﬁcacy

of the recommender system that works on a much

wider range of data coming from different infor-

mation sources.

In the current complex and dynamic labour market,

every organization needs to deploy a continuous and

never-ending learning and formation process in order

to keep its working force and stakeholders always up-

dated and ready to leverage innovative technologies,

new working paradigms and be proactive versus mar-

ket changes. For the same reasons, training institu-

tions need to provide always updated classes and ma-

terials so that students can acquire skills and compe-

tences required by the market.

The challenge for all involved parties in the learn-

ing and formation process is huge. Employees, work-

ers and students (learners) feel the need to be able to

cope with these changes and want to quickly access

to the right, up-to-date and high-quality learning ma-

terial. They would like to be guided through a spe-

ciﬁc learning path customized for their personal back-

ground, ambitions and objectives. They also require

mutual interaction with their peers, teachers and sub-

ject matter experts, in order to exchange knowledge

and experience as well as to feedback regarding the

different contents available.

Subject matter experts and teachers, on the other end,

always need to know which learning material and

contents are mostly requested by users and where to

invest their time. They also seek for high-level com-

parison and advice from their peers like other teach-

ers, domain experts and even learners, in order to cre-

ate learning material with the required quality stan-

dards that match the learners’ needs. To save time

and be quick in updating or producing new courses,

they would like to easily reuse still valid contents and

to be supported in selecting the right course structure.

All organizations and training institutions need to

provide training to their learners in a continuous, per-

vasive way with constantly updated contents. Cost of

training is already relevant and will grow steadily not

leveraging the accumulated knowledge of all learn-

ers or reusing still valid contents for new training ma-

terial or courses. In addition, it becomes mandatory

to easily integrate proprietary contents with free and

publicly available training contents.

CSEDU 2020 - 12th International Conference on Computer Supported Education

112

In this scenario, WhoTeach is a solution then con-

ceived for demanding users, aiming to teach or to

learn in highly-dynamic and complex disciplinary en-

vironments. Speciﬁcally, WhoTeach is the result of

the exploitation of the European project NETT, which

was focused on the conception and design of a so-

cial learning solution for promoting and stimulating

the diffusion of the entrepreneurship knowledge and

mindset in the European countries. In order to con-

tinuously help teachers to conceive, organize and cre-

ate effective and right courses, WhoTeach provides

them with suggestions thanks to the recommender

system. Due to the learning algorithms, the system

identiﬁes relations between the features of the didac-

tic resources that may be relevant, according to the

inquirer’s needs. In this way, teachers and experts

are supported getting intelligent and consistent sug-

gestions, by reducing the high number of available re-

sources. The system intelligence progressively guides

the user in the composition of the course, picking re-

sources in any available format and quickly assem-

bling them to avoid frustration and waste of time.

Moreover, the solution recommends learners the best

training that ﬁts their background and their goals in

terms of skill improvement and career/curriculum de-

velopment in the belonging organization, providing

them with the most updated and customized content

and training material. By the means of the several in-

teraction possibilities available in the social platform,

learners can exchange their experience and feedback

with experts and other users. Teachers and experts

will then receive suggestions on how to improve their

contents based on user comments and which contents

require update.

The learning algorithms allow WhoTeach to dy-

namically choose, organize and update contents,

given the needs, the objectives and the feedback pro-

vided by users. Due to the use of metadata, the con-

tents are homogeneously identiﬁed through a vector

of parameters to represent them. Thereafter, the feed-

back of other users (learners or experts) allow to as-

sociate each composition of vectors to a score. As

a consequence, a dynamical decision tree procedure

leads to where, satisfactory completion of the course

represented by speciﬁc branches of decision trees.

The organization of the learning material is composed

by different knowledge areas related to disciplinary

macro-areas. Contents are divided in resources, mod-

ules and courses. In particular, the resources can vary

in their structure (wiki, discussion forum, eBook, etc)

and format (word, pdf,etc).

In order to highly empower the social learn-

ing processes, especially in relatively new or

highly dynamic learning contexts, the second macro-

component of the platform structure is the so-called

”Knowledge co-creation Social Platform”, based on a

social network providing multiple possibilities of in-

teractions, material exchange and communities cre-

ation, fostered around each disciplinary sector. Es-

pecially oriented to students, the platform helps them

to make them evolve their personal learning experi-

ence, reaching a more collaborative and amazing one.

The platform has standard social network tools (like

blogs, chat, forum, messaging), plus some advanced

features, like the following ones:

• Deﬁnition of Community. Their can be created

around each disciplinary sector, in particular for

teachers, but even for interdisciplinary areas, con-

necting other different communities. Thematic

communities can be freely created by teachers or

experts. All communities are moderated by the

master of the knowledge area. In this way, disci-

plinary aggregations led to communities of prac-

tice and knowledge exchange and improvement.

• Sharing of Didactic Materials. Users can share all

the didactic material, according to their need. In

particular, there is the possibility to share non ofﬁ-

cial material, without waiting experts’ or masters’

approvals. In this way, other feedback can be ac-

quired, thus making the knowledge management

and evolution faster.

• Informal Communication among Users. Private

or public feedback progressively help teachers or

experts in improving their materials. Feedback

are stimulated in the knowledge areas but also

among different areas or disciplines, so as to get

insights for knowledge evolution and disciplinary

innovation.

• Users Proﬁle. Teachers and students have the pos-

sibility to create and edit their own proﬁle: there

is a wide range of possibilities to share personal

experience, academic background, interests, com-

petences and personal or professional ambitions.

Organizations can also interact directly in order to

stimulate talent selection and competences match-

ing.

Groups and Forums. They have several possibil-

ities of interactive features to stimulate coopera-

tion among users. Besides directly sharing ma-

terials, teachers and students can create or share

different kind of contents in a dynamic way (e.g.

tests, links, videos, exercises etc.).

Therefore, this kind of social platform is aimed at

stimulating various types of collaborations among

users. That gives rise to rich, efﬁcient and fruitful

communities of practice to favor course design activ-

ities and allow peer-to-peer learning methodologies.

Attentional Neural Mechanisms for Social Recommendations in Educational Platforms

113

From a technical perspective, the system is based on

a PHP shell piloting and empowering the customiza-

tion of the Moodle platform, serving as a base for a

Content Management System, while Mahara is used

to build the nested social platform. Mahara is a fully-

featured web application to build an electronic port-

folio. A user can create journals, upload ﬁles, embed

social media resources from the web and collaborate

with other users in groups. In Mahara user can con-

trol which items and what information other users see

within their portfolio. The Moodle system was chosen

because of its high diffusion within basically any kind

of training institution and due to the presence of a

wide development community. The platform has been

then integrated with social network features coming

from Mahara, in order to introduce meta-services as

previously described.

3 MAIN CONCEPTS AND

DEFINITIONS

Graphs (annotated with G = (V,E)) are theoretical

objects widely applied to model the complex set of

relationships that typically characterize current net-

works. They consist of a set of “entities ”, V (ver-

tices or nodes), and relationships between them, i.e.

edges, E. In this paper, we use weighted graphs

(graphs whose edges {i, j} ∈ E have assigned weights

label(i, j)), with an associated graph adiacency ma-

trix, A, to indicate whether two vertices are connected

by some edge, i.e., A

i, j

= label(i, j) if {v

} ∈ E.

Moreover, given a vertex v ∈ V , we indicate with

N (v) = { j : {v, j} ∈ E} the neighborhood of the ver-

tex v. We will also indicate with G[A] the graph in-

duced by A. In order to summarize the relationships

between vertices and capture relevant information in

a graph, embedding (i.e., objects transformation to

lower dimensional spaces) is typically applied (Goyal

and Ferrara, 2018). This approach allows to use a rich

set of analytical methods offering to deep neural net-

works the capability of providing different levels of

representation. Embedding can be deﬁned at different

level: for example, at node level, at graph level, or

even through different mathematical strategies. Typ-

ically, the embedding is realized by ﬁtting the (deep)

network’s parameters using standard gradient-based

optimization. The following deﬁnitions can be useful

(Lee et al., 2018).

Deﬁnition 3.1. Given a graph G = (V,E) with V

as the set of vertices and E the set of edges, the

objective of node embedding is to learn a function

f : V → R

such that each vertex i ∈ V is mapped

to a k-dimensional vector,

Deﬁnition 3.2. Given a set of graphs, G, the objective

of graph embedding is to learn a function f : G → R

that maps an input graph G ∈ G to a low dimensional

embedding vector,

4 A GRAPH ATTENTION

MECHANISM FOR

RECCOMENDER SYSTEMS

The principle according to which the attention mech-

anisms play their role is to select the most relevant

information among those available for the neural re-

sponse computation. In other words, “attention” is

essentially a way to non-uniformly weight the contri-

butions of input - or part of it, in order to optimize the

learning process for some speciﬁc task.

There are many way to get this result. In this paper

we consider the case of node embedding–based atten-

tion, as proposed in (Veli

ckovi

c et al., 2017).

Let us consider an user/item relationship matrix A,

and the corresponding weighted graph, G[A] = (V, E),

whose edges are labeled with scores attributed by

users, U ⊆ V , to resources, R ⊆ V , and collected

within A. Given a pair of vertices (u, r), u ∈ U,r ∈ R,

the induced graph representation of A has an edge

between user u and resource r in case that u ap-

plied r and evaluated such a resource with the score

label(u,r). In this way, we have label(u,r) = A

u,r

With the above notation, we conveniently adapt the

deﬁnition of “attention” reported in (Lee et al., 2018)

as follows.

Deﬁnition 4.1. Let A be an user/item relationship

matrix, and G[A] = (U ∪ R, E) its induced weighted

graph with vertices equal to the union of users, U,

and items R, respectively.

Given (u, r), u ∈ U, r ∈ R, an attentional mechanism

for G is a function a : R

× R

→ R which com-

putes coefﬁcients e

(l)

u,r

= a



(l)



across the pairs

of vertices, (u, r), based on their feature representa-

tion

(l)

at level l.

Coefﬁcients e

(l)

u,r

are parameters which act by weight-

ing the relevance of the vertex r’s features to (user) u.

Following (Veli

ckovi

c et al., 2017), we deﬁne a as a

feed-forward neural network with a learnable vector

of parameters (i.e., weights) ~a - updated with the oth-

ers network’s parameters according to standard opti-

mization, and nonlinear LeakyReLU activation func-

tion. In this way, we have

(l)

u,r

= LeakyReLU



(l)

||W

(l)

i

, (1)

CSEDU 2020 - 12th International Conference on Computer Supported Education

114

where W is a learnable parameter matrix and

(l)

||W

(l)

is the concatenation of the embed-

ded representation for the vertices u,r.

The coefﬁcients e

(l)

u,r

can be normalized over all ele-

ments in the neighborood of u. For instance, we can

apply the softmax function to obtain the following ex-

pression.

(l)

u,r

exp(e

(l)

u,r

)

∑

k∈N (u)

exp(e

(l)

u,k

)

At this point, the role of (coefﬁcients) α

(l)

u,r

becomes

fundamental for our extension. Let us consider the

induced graph G[A] = (U ∪ R, E), and focus on some

user (vertex) u ∈ U. When only resources (items)

around u are considered, we can use the normalized

(attention) coefﬁcients α

(l)

u,r

to compute a combination

of (the embedded resources)

(l)

in N (u) as follows.

(l+1)

= σ



∑

r∈N (u),r∈R

(l)

u,r

(l)



(2)

where σ is non linear vector-valued function (sig-

moid). With this formulation, Eq. 2 provides the

next level embedding for user u scaled by the atten-

tion coefﬁcients which, in turn, can be interpreted as

the relevance of the resources used by (user) u (i.e.,

resources in the neighborood of u). Similarly to Eq.

2, the following quantity can be interpreted as the em-

bedded representation for the resource r scaled by the

attention coefﬁcients which weight users (representa-

tion) who applied r (i.e., users in the neighborood of

r).

(l+1)

= σ



∑

u∈N (r),u∈U

(l)

u,r

(l)



(3)

In this way, the “GAT layer” returns for each

pair (u,r) ∈ U × R the embedded representation

(

(l+1)

). In our experiments we will consider

only one level of embedding, i.e., l = 1.

4.1 A Stacked Architecture for

Social-based Recommendations

The attentional mechanism described in this paper

was applied as a “base” layer (Module A) for the

stacked architecture reported in Fig. 1. Two outputs

are provided:

(l+1)

i.e., the embedded representation

for user-u’s score, and

(l+1)

, the embedded represen-

tation for resource-r’s scores, respectively. The fol-

lowing details summarize this layer.

• Architecture: described in the previous para-

graph.

Input: Given the user/item matrix, A = (s

i, j

which contains the score, s

i, j

, for each user i

and resource (item) j, a training set of examples

T = {((~u

,~r

),s

i, j

) : 1 ≤ s

i, j

≤ 5} was obtained by

composing the vector of scores, ~u

(provided by

user i for each available resource), and the vector,

(scores attributed by all users to the resource j).

Output: For each i and j, the embedded user-i’s

scores

(l+1)

and item- j’ s scores

(l+1)

Two embedding,

(l+1)

are then passed and

combined through feed forward levels (FFL) in or-

der to obtain, using a ﬁnal sigmod-based activation,

the score predicted for the user/item (input) pair (i, j).

The whole model is trained with MSE loss and SGD

(stochastic gradient descent) optimizer. In particu-

lar the following general architecture (Fig. 1a) was

stacked on the top of the attention layer.

• Stacked layer (Module B in Fig. 1a).

Input:

(l+1)

Output: For each i and j, the predicted score, ˆs

i, j

for user i when choosing (resource) j.

FFL, Sub-mod. n.1:

(l+1)

as input + Dense

layer with ReLU Activation function + h

as em-

bedded output representation for user-i’s scores.

FFL, Sub-mod. n.2:

(l+1)

as input + Dense

layer with ReLU Activation function + h

embedded output representation for resource- j’s

scores.

Operator Layer:

(l+1)

and

(l+1)

are combined

(through a speciﬁc operator) to obtain the vector

(

(l+1)

Dense Layer: This ﬁnal layer uses an output

sigmoid activation. The value ˆs

i, j

assumed by

the sigmoid function is then interpreted as output

score for user i and resource j.

Note that the architecture described above shows a

general structure, and can provide different models

according to the type of operation applied by the con-

sidered “operator layer”.

5 NUMERICAL EXPERIMENTS

Numerical experiments use an homogeneous set of

data whose characteristics combine well with the re-

quirements of the WhoTeach platform. These data

come from the “Goodbooks” data-set (https://www.

kaggle.com/zygmunt/goodbooks-10k), a large collec-

tion reporting up to 10000 books and 1000000 rat-

ings assigned by 53400 readers. In particular, numer-

ical ratings, ranging from “1” to “5”, are given by

Attentional Neural Mechanisms for Social Recommendations in Educational Platforms

115

(a)

(b)

Figure 1: System Architecture. A feed-forward based net-

work (Module B) is stacked on the top of the GAT layer

(Module A). The embedding of user-u’s scores

(l+1)

, and

item-r’s scores

(l+1)

, are passed and combined (“opera-

tor layer” in Mod. B) with further embedding (FFL Sub-

modules) in order to provide through a sigmod-based acti-

vation (output of “Dense layer”) the predicted ﬁnal sugges-

tion.

users (readers) for each resource type, in this case,

different sort of book. Numerical experiments are

planed to evaluate the capability of the attentional-

based models to reduce error (loss function) between

the reported and predicted preference scores. To pro-

vide robust estimation, we sub-sampled the data us-

ing cross-validation. The models described in this

paper was implemented using the Pytorch library

(https://pytorch.org/), and then executed using differ-

ent parameters for early stopping and learning rate, on

COLAB (https://colab.research.google.com/). In this

work in progress the attention-based model with con-

catenation operator in the stacked layer (see Fig. 1)

was compared with the following alternative models.

Performances were averaged on the number of folds

(10 cross-validation).

1. Dot product model.

Input: Training set T = {((u

),s

i, j

) : 1 ≤

i, j

≤ 5} as described previously for the atten-

tional based architectures.

Output: for each i, j, the score ˆs

i, j

recommended

for user i and resource j.

Loss Function: MSE; Optimizer: SGD.

Architecture: similar to the stacked architecture

(with no attention). A dot product operation is

Table 1: MSE comparison: Attention is applied with the

concatenation operator at the stacked layer. Hadamard uses

element-wise product between two vectors at the operator

layer.

Model Attention Concatenate Multiply Hadamard

MSE 0.0389 0.0439 0.0437 0.0436

applied to “aggregate” the embedded representa-

tions of ~u

and ~r

. No learnable parameters are

considered.

2. Element-wise product model (Hadamard product

model).

Similar to the previous case (dot product model)

but with an element wise product operation

(Hadamard product between vectors) computed

by the Operator Layer. The result of the element-

wise operation is is passed to the ﬁnal dense layer.

3. Concatenation model.

Similar to the dot product model but with a con-

catenation operation computed by the Operator

Layer. In this case, the embedded representations

of~u

and~r

are concatenated in a new latent vector

and ﬁnally passed to a dense layer.

Preliminary results are reported in Tab. 1. A gen-

eral better tendency to reduce the MSE loss is ob-

served when attention layer with concatenation is ap-

plied as a base module for the considered stacked

layer.

6 CONCLUSIONS

Online social spaces are rich sources of information

for modern recommender systems. However, the suc-

cess of reliable recommendations is related to both the

capability of a framework to capture the social con-

tent and the availability of an effective information in

the considered online social space. The work reported

in this paper has focused on a recent ”mechanism”

formulation for learning on graphs with “attention”

(Veli

ckovi

c et al., 2017). In particular, the proposed

architecture intends to beneﬁt from exploiting (with

“attention” weights) the graph’s task-relevant part in

order to provide reliable social recomendations. As

recently reported (Seo et al., 2017; Lu et al., 2018),

these mechanisms constitute challange solutions to

provide users with effective (social) justiﬁcations for

the suggestions that modern recommendation systems

are able to offer. Our research will follow now this tar-

get by adding to the social engine of our “WhoTeach”

platform textual suggestions based on the computed

“attention weights” as deﬁned in Sec. 4. Numerical

results obtained in the experiments are encouraging.

CSEDU 2020 - 12th International Conference on Computer Supported Education

116

The proposed framework outperform feed forward-

based networks when the attention layer is applied.

Our future research in this context will also consider

additional architectures and data for further evalua-

tions and more general conclusions.

REFERENCES

Bahdanau, D., Cho, K., and Bengio, Y. (2014). Neural ma-

chine translation by jointly learning to align and trans-

late. arXiv preprint arXiv:1409.0473.

Bonhard, P. and Sasse, M. A. (2006). ’knowing me, know-

ing you’—using proﬁles and social networking to im-

prove recommender systems. BT Technology Journal,

24(3):84–98.

Chen, X., Xu, H., Zhang, Y., Tang, J., Cao, Y., Qin, Z., and

Zha, H. (2018). Sequential recommendation with user

memory networks. In Proceedings of the Eleventh

ACM International Conference on Web Search and

Data Mining, WSDM ’18, pages 108–116, New York,

NY, USA. ACM.

Dondi, R., Mauri, G., and Zoppis, I. (2016). Clique editing

to support case versus control discrimination. In Intel-

ligent Decision Technologies, pages 27–36. Springer.

Dondi, R., Mauri, G., and Zoppis, I. (2019). On the

tractability of ﬁnding disjoint clubs in a network. The-

oretical Computer Science.

Frasconi, P., Gori, M., and Sperduti, A. (1998). A general

framework for adaptive processing of data structures.

IEEE transactions on Neural Networks, 9(5):768–

786.

Gori, M., Monfardini, G., and Scarselli, F. (2005). A new

model for learning in graph domains. In Proceedings.

2005 IEEE International Joint Conference on Neural

Networks, 2005., volume 2, pages 729–734. IEEE.

Goyal, P. and Ferrara, E. (2018). Graph embedding tech-

niques, applications, and performance: A survey.

Knowledge-Based Systems, 151:78–94.

Gupta, P., Goel, A., Lin, J., Sharma, A., Wang, D., and

Zadeh, R. (2013). Wtf: The who to follow service at

twitter. In Proceedings of the 22nd international con-

ference on World Wide Web, pages 505–514. ACM.

Lee, J. B., Rossi, R. A., Kim, S., Ahmed, N. K., and Koh, E.

(2018). Attention models in graphs: A survey. arXiv

preprint arXiv:1807.07984.

Li, C., Quan, C., Peng, L., Qi, Y., Deng, Y., and Wu, L.

(2019). A capsule network for recommendation and

explaining what you like and dislike.

Lu, Y., Dong, R., and Smyth, B. (2018). Coevolutionary

recommendation model: Mutual learning between rat-

ings and reviews. In Proceedings of the 2018 World

Wide Web Conference, pages 773–782.

Scarselli, F., Gori, M., Tsoi, A. C., Hagenbuchner, M.,

and Monfardini, G. (2008). The graph neural net-

work model. IEEE Transactions on Neural Networks,

20(1):61–80.

Schafer, J. B., Frankowski, D., Herlocker, J., and Sen, S.

(2007). Collaborative ﬁltering recommender systems.

In The adaptive web, pages 291–324. Springer.

Seo, S., Huang, J., Yang, H., and Liu, Y. (2017). Inter-

pretable convolutional neural networks with dual local

and global attention for review rating prediction. In

Proceedings of the eleventh ACM conference on rec-

ommender systems, pages 297–305.

Sperduti, A. and Starita, A. (1997). Supervised neural net-

works for the classiﬁcation of structures. IEEE Trans-

actions on Neural Networks, 8(3):714–735.

Veli

ckovi

c, P., Cucurull, G., Casanova, A., Romero, A., Lio,

P., and Bengio, Y. (2017). Graph attention networks.

arXiv preprint arXiv:1710.10903.

Zhang, Y. and Chen, X. (2018). Explainable recommenda-

tion: A survey and new perspectives.

Zhou, X., Xu, Y., Li, Y., Josang, A., and Cox, C. (2012).

The state-of-the-art in personalized recommender sys-

tems for social networking. Artiﬁcial Intelligence Re-

view, 37(2):119–132.

Zoppis, I., Dondi, R., Manzoni, S., Mauri, G., Marconi, L.,

and Epifania, F. (2019a). Optimized social explana-

tion for educational platforms. In Int. Conf. on Com-

puter Supported Education, volume 1, pages 85–91.

Zoppis, I., Manzoni, S., and Mauri, G. (2019b). A com-

putational model for promoting targeted communica-

tion and supplying social explainable recommenda-

tions. In 2019 IEEE 32nd International Symposium

on Computer-Based Medical Systems (CBMS), pages

429–434. IEEE.

Attentional Neural Mechanisms for Social Recommendations in Educational Platforms

117