Learning and training processes are starting to be a ected by the di usion of Arti cial Intelligence (AI) techniques and methods. AI can be variously exploited for supporting education, though especially deep learning (DL) models are normally su ering from some degree of opacity and lack of interpretability. Explainable AI (XAI) is aimed at creating a set of new AI techniques able to improve their output or decisions with more transparency and interpretability. Deep attentional mechanisms proved to be particularly e ective for identifying relevant communities and relationships in any given input network that can be exploited with the aim of improving useful information to interpret the suggested decision process. In this paper we provide the rst stages of our ongoing research project, aimed at signi cantly empowering the recommender system of the educational platform "WhoTeach" by means of explainability, to help teachers or experts to create and manage highquality courses for personalized learning. The presented model is actually our rst tentative to start to include explainability in the system. As shown, the model has strong potentialities to provide relevant recommendations. Moreover, it allows the possibility to implement e ective techniques to completely reach explainability 3.
Nowadays, learning and training processes are starting to be a ected by the
diffusion of AI techniques and methods [
In this context, explainable AI in the eld of Recommender Systems (XRS) is
aimed at providing intuitive explanations for the suggestions and
recommendations given by the algorithms [
Speci cally, in literature explainable attentional models have been used in domains ranging from medical care [24] [25] to e-commerce and online purchase [26]. It is worth to notice that attention weights can have a role in starting to fostering explainability [27].
In this article, we provide the rst stages of our ongoing research project, aimed at signi cantly empowering the RS of our educational platform "WhoTeach" [29] by the means of explainability. Speci cally, we report our current positioning in the state of the art with the proposed model to extend the social engine of \WhoTeach" with a graph attentional mechanism aiming to provide social recommendations for the design of new didactic programs and courses. The presented model allows us to start to include explainability in the system.
We have started to de ne our positioning in the state-of-the-art according
to three dimensions studied in the XAI literature [
In particular:
{ Display style: in order to optimize the user experience we are working to
de ne the way explanations will be provided, as the di erent possibilities in
literature show [
From the study of the state of the art, we then strive to inscribe our current work and its future steps in the XRS literature, so as to de ne our present positioning and prepare for future work and stages towards explainability. 2
WhoTeach (WT) is a complete digital learning platform for supporting heterogenous learning ecosystems in their processes and activities, due to its numerous synchronous and asynchronous features and functionalities. WT is aimed at promoting the development of customized learning and training paths by aggregating and disseminating knowledge created and updated by experts. The platform is conceived as a Social Intelligent Learning Management System (SILMS) and it is structured around three components: 1. The Recommender System (RS), to help experts and teachers to quickly and e ectively assemble high-quality contents into courses: thanks to an intelligent analysis of available material, it is aimed at suggesting teachers the best resources to include, in any format, according to teachers' needs or requirements. 2. The "Knowledge co-creation Social Platform", which is a technological infrastructure based on an integrated and highly interactive social network, endowed with many features to share information, thematic groups and discussion forums. 3. The content's repository where to upload contents from any course or training material, either proprietary or open. This serves as a basis for both the recommender system to elaborate materials and also users who want to create personalized courses. 3
A graph (annotated with G = (V; E)) is a theoretical object widely applied to model the complex set of relationships that typically characterize current networks. This object consists of a set of \entities " (vertices or nodes), V , and relationships between them, i.e. edges, E. In this paper, we use attributed graphs, i.e., particular graphs where each vertex v 2 V is labeled with a set of attribute values. Moreover, given a vertex v 2 V , we indicate with N (v) = fu : fv; ug 2 Eg the neighborhood of the vertex v.
Given a graph G, we use the corresponding adjacency matrix A to indicate whether two vertices vi; vj of G are connected by an edge, i.e., (A)i;j = 1, if fvi; vj g 2 E.
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 [23]. This approach allows to use a rich set of
analytical methods, o ering to deep neural networks the capability of
providing di erent levels of representation. Embedding can be performed at di erent
level: for example, at the node level, at the graph level, or even through
different mathematical strategies. Typically, the embedding is realized by tting
the (deep) network's parameters using standard gradient-based optimization. In
particular, the following de nitions can be useful [
De nition 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 ! Rk such that each vertex i 2 V is mapped to a k-dimensional vector, h. De nition 2. Given a set of graphs, G, the objective of graph embedding is to learn a function f : G ! Rk that maps an input graph G 2 G to a low dimensional embedding vector, h. 4
In our application, we use the attentional-based node embedding proposed in
[
De nition 3. Let A be an user/item relationship matrix, G[A] = (V; E) the corresponding weighted graph, and V = fU; Rg the set of users U and items R, respectively. Given a pair of vertices (u; r); u 2 U; r 2 R, an attentional emue;rch=anaismh(ulf)o; rh(rGl) isacarofsusncthtieonpaair:s RofnverRticnes!,uR;r,wbhaiscehd coonmtphuetiers fceoaeturecierenptsresentation h(ul); h(rl) at level l.
Coe cients eu;r are considered as the importance of the vertex r's features to (user) u.
Following [
The coe cients eu;r can be normalized using, e.g., the softmax function u(l;)r =
exp(e(ul;)r)
Pk2N (u) exp(e(ul;)k) : The mechanisms parameters, a, are then updated with the others network's parameters accordingly to typical optimization algorithms. When only resources (items) around u are considered, the normalized (attention) coe cients u;r can be used to compute a combination of the resources h(rl) in N (u) as follows h(ul+1) =
X r2N (u);r2R u(l;)rW(l)h(l) r where is non linear vector-valued function (sigmoid). With this formulation, Eq. 2 provides the next level embedding for user u scaled by the attention scores which, in turn, can be interpreted as the relevance of the resources used by the user u. Similarly to Eq. 2, the following quantity can be interpreted as the user scores who applied, in particular, the resource r. (2) (3) h(l+1) = r
X u2N (r);u2U u(l;)rW(l)h(ul) In this way, the \GAT layer" returns for each pair (u; r) 2 U R the embedded representation (h(ul+1); h(rl+1)). In our experiments we will consider only one level of embedding, i.e., l = 1.
Therefore, as previously described in the section 3, we introduce a novel kind of information representation h(ul) for users and h(rl) resources, allowing us to visualize either the user u or the resource r as the main element according to its neighborhood. Nevertheless, this representation is still not able to explain and justify the recommendations given to a speci c user. Indeed, it provides the starting point to apply the attention mechanism, which introduces the possibility to give a weight e(ul;)r to the most relevant information encoded in the embedded representation for both the user hu (l+1). Then, (l+1) and the resource hr the attention weights e(ul;)r permit to improve the model performances, reducing the error for the recommendations.
Above all, they foster the possibility to explain why a given resource r is recommended to a speci c user u. In particular, this approach for computing the attention weights e(ul;)r is also applied in other works related to collaborative ltering RS [32]. Other works explore di erent display styles, as visual explanations [31]. In conclusion, the ability to highlight the most useful information to realize the recommendations allowed us to start to introduce explainability in the system.
Here we report a short review of the numerical experiments described in [29]. The experiments use an homogeneous set of data whose characteristics combine well with the requirements of the WhoTeach platform. These data come from the \Goodbooks" data-set (https://www.kaggle.com/zygmunt/goodbooks-10k), a large collection reporting up to 10K books and 1000000 ratings (from \1" to \5") assigned by 53400 readers.
The experiments aim to evaluate the capability of the attentional-based models to reduce error (loss function) between the reported and predicted preference scores.
The models was implemented using the Pytorch library (https://pytorch.org/), and then executed using di erent hyperparameters.At the present stage, the attention-based model was compared with alternative models: dot product model, element-wise product model (Hadamard product model), concatenation model. Performances were averaged on the number of folds (10 cross-validation).
A general better tendency to reduce the MSE loss is observed when attention layer with concatenation is applied as a base module for the considered stacked layer. 6
In this work we have reported our work in progress for providing \WhoTeach" with an explainable recommender system, aimed to signi cantly empower its ability to help teachers or experts to create high-quality courses. It is totally clear that further improvements of the present XRS could signi cantly help users to better understand the reason why speci c items are recommended. At the present stage, we started to propose a model based on the attentional mechanisms, which allows to justify the chosen recommendations provided by the model by the means of the attention weights. This model is speci cally focused on exploiting social information for educational services, thus extending the social engine of our educational platform \WhoTeach" to reinforce the AI engine. Finally, we have reported our present and further positioning in the state of the art of XAI, showing the potentialities of the present model and the next steps in our work. One of the most important further step will be the optimization of the computational complexity for the computation of the attention weights. Moreover, we will strive to improve the display style of the explanations provided to users, in order to consequently improve the user experience.