A Graph Neural Network For Fuzzy Twitter Graphs Georgios Drakopoulos1 , Eleanna Kafeza2 , Phivos Mylonas1 and Spyros Sioutas3 1 Department of Informatics, Ionian University, Tsirigoti Sq. 7, Kerkyra 49100, Hellas 1 College of Technological Innovation, Dubai Academic City, E-L1-108, UAE 3 Computer Engineering and Informatics Department, University of Patras, Patras 26504, Hellas Abstract Social graphs abound with information which can be harnessed for numerous behavioral purposes including online political campaigns, digital marketing operations such as brand loyalty assessment and opinion mining, and determining public sentiment regarding an event. In such scenarios the efficiency of the deployed methods depends critically on three factors, namely the account behavioral model, the social graph topology, and the nature of the information collected. A prime example is Twitter which is especially known for the lively activity and the intense conversations. Here an extensible computational methodology is proposed based on a graph neural network operating on an edge fuzzy graph constructed by a combination of structural, functional, and emotional Twitter attributes. These graphs constitute a strong algorithmic cornerstone for engineering cases where a properly formulated potential or uncertainty functional is linked to each edge. Starting from the ground truth in each individual vertex, the graph neural network progressively computes in an unsupervised manner a global graph state which can in turn be subject to further processing. The results, obtained using as a benchmark a recent similar graph neural network architecture along with two Twitter graphs, are promising. Keywords Fuzzy graphs, graph mining, graph neural networks, behavioral analytics, emotional polarity, Twitter 1. Introduction graph neural network (GNN) architecture. GNNs con- stitute a class of unsupervised neural networks where Graph mining is an integral part of the interconnected each vertex, representing a processing node, starts with era since it lays the groundwork for numerous applica- a local ground truth information vector and iteratively tions across a wide array of financial and technological a global status is derived based on the fundamental fact fields including among others social network analysis, that graphs contain inherently higher order information database query optimization, graph signal processing in a distributed manner. The resulting graph global state (GSP), supply chain and logistics networks, and brain cir- can be subsequently further processed in order to derive cuit analysis. In this context modeling a graph in terms global properties such as community discovery. of vertices, connectivity patterns, and associated features The primary research objective of this conference pa- is tantamount to data model selection. Edge fuzzy graphs per is the development of a GNN architecture designed extend classical graphs as probabilities drawn from a sin- for edge fuzzy Twitter graphs constructed from incor- gle distribution which may well have unknown param- porating structural, functional, and behavioral features. eters to be estimated. Said distribution is closely linked The proposed methodology can be inherently extended to to the semantics and functional nature of the underlying other possible attribute types, making it thus appropriate graph. For instance, in a transportation network edge ex- for mining graphs originating from social media or evolv- istence probabilities can show how likely a specific road ing computational ecosystems for that matter. This work is to be blocked from snow in winter months, whereas differentiates itself from previous ones in two aspects, in a computer network they may model the chance of a namely the fusion of various heterogeneous attributes virus being propagated along a given link. and the induced edge fuzzy topology. In order to compute an estimation of the global graph The remaining of this work is structured as follows. In state which allows not only a higher level overview but section 2 the recent scientific literature regarding GNNs, also subsequent processing, in this work will be used a graph mining, and computational behavioral science is briefly reviewed. The proposed methodology along with CIKM’21: 30th ACM International Conference on Information and the relevant intuition are given in 3. The results obtained Knowledge Management, November 01–05, 2021, Virtual Event, QLD, from the experiments are the focus of section 4. Future Australia $ c16drak@ionio.gr (G. Drakopoulos); eleana.kafeza@zu.ac.ae research directions are given in 5. Technical acronyms (E. Kafeza); fmylonas@ionian.gr (P. Mylonas); are defined the first time they are encountered in the sioutas@ceid.upatras.gr (S. Sioutas) text. Finally, the notation of this conference paper is  0000-0002-0975-1877 (G. Drakopoulos); 0000-0001-9565-2375 summarized in table 1. (E. Kafeza); 0000-0002-6916-3129 (P. Mylonas) © 2021 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). CEUR Workshop Proceedings http://ceur-ws.org ISSN 1613-0073 CEUR Workshop Proceedings (CEUR-WS.org) 1 Georgios Drakopoulos et al. CEUR Workshop Proceedings 1–7 Table 1 Notation of this conference paper. Symbol Meaning First in △ = Definition or equality by definition Eq. (1) {𝑠1 , . . . , 𝑠𝑛 } Set with elements 𝑠1 , . . . , 𝑠𝑛 Eq. (2) (𝑡1 , . . . , 𝑡𝑛 ) Tuple with elements 𝑡1 , . . . , 𝑡𝑛 Eq. (1) |𝑆| Set cardinality functional Eq. (3) prob {Ω} Probability of event Ω occurring Eq. (4) 2. Previous Work 3.1. Fundamental concepts GNNs operate on irregular domains expressing relation- In order to describe the proposed architecture a few basic ships. Heterogeneous GNN architectures are examined concepts must be first revised or defined. First the class in [1] and representative GNNs designed to complete ver- of edge fuzzy graphs is introduced in definition 1. satile tasks in [2]. Edge labeling is proposed in [3] in the Definition 1 (Edge fuzzy graph). An edge fuzzy context of few-short learning for GNNs. The technique graph is a combinatorial object represented by the ordered of aggregated neural path in conjunction with machine triplet shown in equation (1). learning (ML) tasks is described in [4]. The emotional coherency of Twitter graphs with GNNs is explored in △ 𝐺 = (𝑉, 𝐸, ℎ) (1) [5], whereas in [6] are given guidelines for social recom- mendation based on GNNs. The elements in (1) have the following meaning: Graph mining is a mainstay of current ML [7]. In a graph signal processing (GSP) context adjacency matrices • The vertex set 𝑉 . In the context of this work are considered as two-dimensional signals and signal pro- each vertex corresponds to a single Twitter account cessing techniques are then employed to extract patterns through a bijection. of interest [8]. An overview of the connections to deep • The set of fuzzy edges 𝐸 where 𝐸 ⊆ 𝑉 × 𝑉 . The learning are given in [9]. In [10] a tensor stack network connectivity patterns therein reflect the underlying (TSN) is trained to estimate the topological correlation graph dynamics. of graph pairs compressed with the two-dimensional dis- • The functional ℎ : 𝐸 → [0, 1] maps each edge to a crete cosine transform (DCT2), while the same architec- probability drawn from a single distribution. These ture evaluates graph resiliency in [11]. Flow-based GSP result from graph semantics and functionality. is examined in [12]. The basic operations of GSP such as In the general case the digital account behavior for any shifting and sampling are defined in [13]. A graph ver- online social network is given in definition 2. sion of the LMS adaptive filtering algorithm is presented in [14]. A versatile and space efficient data structure for Definition 2 (Account behavior). The online behavior persistent graphs is described in [15]. of an account consists of the total peer interaction over all Behavioral attributes have recently emerged as an inte- possible ways offered by the given social medium. gral part of many recent computational systems [16]. The The above definition can be readily extended in the case connection between behavioral systems and data driven two or more accounts are connected over multiple social analysis is explored in [17]. Digital trust is a paramount media, expanding thus the interaction potential. How- factor for recruiting candidates from LinkedIn [18]. Clus- ever, this is outside the scope of this work. tering fMRI images with tensor distances for emotion In this work the online behavior of Twitter accounts recognition [19], while gamification strategies are ex- has three distinct components, namely the follow rela- plored in [20]. An overview of behavioral systems is tionships, retweet patterns, and emotional polarity with given in [21]. respect to a reference hashtag set. The intuition behind their selection is as follows: 3. Proposed Architecture • The follow relationships capture the structural In this section the proposed GNN architecture as well as aspect of the Twitter graph since they constitute the notions underlying it are described. the core of its edges. • The retweet patterns are an integral part of the functionality taking place bridging accounts in a different way. 2 Georgios Drakopoulos et al. CEUR Workshop Proceedings 1–7 Account the graph with a certain probability which in the general behavior case depends on an attribute set. The latter is frequently strictly local or a function of a small neighborhood and rarely global since updating such a set is costly and prone to dependency bottlenecks. In the context of this con- Follow Retweet Emotional ference paper the probability 𝑝𝑖,𝑗 for edge 𝑒𝑖,𝑗 between patterns patterns polarity vertices 𝑣𝑖 and 𝑣𝑗 is computed as in equation (4): △ 𝑟𝑖,𝑗 Figure 1: Behavioral model. prob {𝑒𝑖,𝑗 } = 𝑝𝑖,𝑗 = 𝑤𝑓 𝐹𝑖,𝑗 + 𝑤𝑟 + 𝑤𝑐 𝑐𝑖,𝑗 (4) 𝑅 In (4) three factors are taken into consideration: • The emotional coherency is a factor evaluating • Whether there is a directed follow link from the the similarity of sentiments towards selected top- 𝑖-th account to the 𝑗-th one denoted by the binary ics expressed as hashtags. indicator 𝐹𝑖,𝑗 . • The ratio of the number 𝑟𝑖,𝑗 of retweets of the The above are also shown in figure 1. 𝑖-th account coming from the 𝑗-th one to the total In order to model the behavioral aspects of the Twitter retweets 𝑅 in the graph. accounts, a set of the most common hashtags from each • The signed correlation factor 𝑐𝑖,𝑗 expressing the graph is selected. The inspiration for the selection of such emotional coherency of the 𝑖-th and 𝑗-th accounts a set is the concept of node cover. Definition 3 clarifies it. with respect to the reference hashtag set. Definition 3 (Reference hashtag set). Let 𝐻0 be the The sentiment 𝑙[𝑡] of the 𝑖-th account during iteration set of hashtags in a Twitter graph 𝑇 . A hashtag ℎ ∈ 𝐻0 is 𝑡 consists of a vector containing an emotional polarity also belongs to the reference hashtag set if and only if the score, namely the percentage of the how positive, neutral, accounts who have used ℎ constitute a vertex cover for 𝑇 . of negative the 𝑖-th account feels towards the as shown Let 𝐻 be the set of hashtags satisfying definition 3. in (5). Initially the ground truth vector of the 𝑖-th account is the respective average percentage of positive, neutral, △ or negatively charged words in the tweets containing at 𝐻 = {ℎ1 , . . . , ℎ𝑝 } (2) least one hashtag from the reference set. The ratio of the cardinality of 𝐻 to that of 𝐻0 can be [𝑡] △ ]︀𝑇 (5) [︀ taken as a measure of the important information existing 𝑙𝑖 = 𝑛𝑝,𝑖 𝑛𝑛,𝑖 𝑛𝑔,𝑖 in the underlying Twitter graph as in (3): Given the iteration-dependent value of 𝑙[𝑡] , the value △ |𝐻| of the correlation factor 𝑐𝑖,𝑗 should be computed dur- 𝜌 = (3) |𝐻0 | ing each iteration as shown in (6). It should be high- lighted that 𝑐𝑖,𝑗 is the only term of (4) which is signed, Using the notion of the vertex set to find popular hash- thereby reinforcing or weakening the strength between tags has the following advantages: two accounts depending on whether they have similar sentiments towards the reference hashtag set. • Selecting hashtags ℎ does not depend on any hy- perparameters or on any thresholds whatsoever. ∑︀3 (︁ [𝑡] )︁ ∑︀ 3 (︁ [𝑡] )︁ 𝑘=1 𝑙𝑖 [𝑘] − 1/3 𝑘=1 𝑙𝑗 [𝑘] − 1/3 • The widespread use of hashtag ℎ is a clear indi- √︂ cation of its popularity. ∑︀3 (︁ [𝑡] )︁2 √︂∑︀ 3 (︁ [𝑡] )︁2 𝑙 [𝑘] − 1/3 𝑙𝑗 [𝑘] − 1/3 • Although vertex cover is an NP-hard problem, 𝑘=1 𝑖 𝑘=1 approximation algorithms for it exist. (6) The weights of the linear combination in (4) encode However, it should be noted that for larger benchmark the sign and relative strength of each factor contribu- graphs or for dynamic ones alternative criteria should be tion, namely how much each factor participates to the sought in order to avoid the overwhelming complexity edge probability existence and whether such participa- of determining a vertex cover. tion reinforces or weakens said probability respectively. Moreover, they ensure the numerical stability of 𝑝𝑖,𝑗 . 3.2. Architecture Intuitively speaking, equation (4) is a linear estima- tor of the true edge existence probability. The weights The proposed GNN architecture relies on the fundamen- 𝑤𝑓 , 𝑤𝑟 , and 𝑤𝑐 express the relative contribution of each tal fact that edges are fuzzy, namely that they belong to term and in our experiments follow the semantic strength 3 Georgios Drakopoulos et al. CEUR Workshop Proceedings 1–7 of the respective factor. This means that 𝑤𝑓 is higher The hyperparameter 𝛽0 scales input to a practical do- since the follow denotes a high degree of coherency be- main for the sigmoid function 𝜙 (·), 𝛿𝑖,𝑘 is the weight of tween the two accounts. Along a similar line of reason- the edge, and ∆ is the sum of the edge weights of the in- ing, frequent retweets between two accounts indicate a bound neighbors. In (11) the sigmoid function is defined somewhat strong connection between them. Moreover, as in (12) which is differentiable and smooth everywhere. a consistent emotional coherency between two accounts may well suggest a behavioral link between them. △ 1 𝜙 (𝑠; 𝜎0 ) = (12) Additionally the weight 𝛿𝑖,𝑗 assigned to each edge is a 1 + exp (−𝜎0 𝑠) function of the strength of the corresponding edge. The derivative of the sigmoid function is given in (13). △ 𝛿𝑖,𝑗 = 𝑓 (𝑝𝑖,𝑗 ) (7) 𝜕𝜙 (𝑠; 𝜎0 ) = 𝜎0 𝜙 (𝑠; 𝜎0 ) 𝜙 (−𝑠; 𝜎0 ) 𝜕𝑠 The weight function 𝑓 (·) of (7) is the same for each = 𝜎0 𝜙 (𝑠; 𝜎0 ) (1 − 𝜙 (𝑠; 𝜎0 )) (13) edge and it is directly or at least indirectly linked to the semantics of the underlying graph. One of the most The last form of (13) comes from the fundamental common options is that shown in (8). property of the sigmoid function described in (14) below: △ 1 (14) 𝛿𝑖,𝑗 = (8) 𝜙 (𝑠; 𝜎0 ) + 𝜙 (−𝑠; 𝜎0 ) = 1 𝑝𝑖,𝑗 The preceding properties ensure that 𝜙 (·) is smooth However, the weight selection of (8) has the disadvan- enough to prevent divergence in most cases for a broad tage of being almost singular close to zero, generating spectrum of distributions. thus excessive weight values. A viable alternative is the inverse linear weight function of (9). △ 1 4. Results 𝛿𝑖,𝑗 = (9) 1 + 𝑝𝑖,𝑗 The results of the proposed GNN methodology are pre- sented in this section along with intuition about them. Another option for the weight function is that the in- They are divided to four groups, one for each possible verse square function of equation (10). The latter typically combination of benchmark graph (1821 / US2020) and expresses a potential function in various applications. weight function (inverse linear / inverse square). △ 1 𝛿𝑖,𝑗 = (10) 1 + 𝑝2𝑖,𝑗 4.1. Dataset In figure 2 the weight functions of (9) and (10) are The two benchmark graphs used in the experiments were shown for their entire range. It can be immediately taken from [5]. They represent two characteristic cases inferred they are strictly decreasing and everywhere of social graphs, namely one with a relative quiet and smooth, expressing the fact that the more likely is an coherent one (1821) and one containing heated conversa- edge to belong to the graph, the easier to cross it. tions and a considerable degree of dissonance (US2020). At the core of the proposed GNN architecture is the The Twitter sampling interval was 8/2020-10/2020. [𝑡] update mechanism of (11). For the 𝑖-th vertex the 𝑙𝑖 is computed as in (11). Therein the index 𝑗 ranges over 4.2. Number of iterations all inbound neighbors of the 𝑖-th vertex and thus it de- In table 3 the parameters used in the experimental setup pends on local connectivity patterns. However, since the of this work are shown. This allows for the easy explo- state vector of its neighbors depends on recursively on ration of the parameter space. Observe that the actual val- that of its own vectors, this mechanism is essentially a ues of these parameters are in accordance of the strength higher order status computation. During an update it of the respective factor. may be possible that certain neighbors may have already Table 4 contains the normalized number of iterations had their own state vectors updated, whereas others not. as a function of the hyperparameter 𝛽0 of (11) for the Thus, the iteration indicator * will be used. This process two benchmark graphs of table 2 and for the two possible terminates when the state vectors remain unchanged weight functions shown in equations (9) and (10). Nor- under a threshold of 𝜂0 for three consecutive iterations. malization takes place per graph and per weight function (︃ )︃ in order to show the comparative effect of 𝛽0 in each 𝛽0 [𝑡−1] 𝛽0 ∑︁ 𝛿𝑖,𝑗 [*] 𝑙 [𝑡+1] = 𝜙 𝑙 + 𝑙 (11) case. In order to demonstrate the effect of the emotional 2 2 𝑗 ∆ 𝑗 attribute 𝑐𝑖,𝑗 of (4) on the convergence rate the same 4 Georgios Drakopoulos et al. CEUR Workshop Proceedings 1–7 Weight functions vs edge probability 1 inv.linear inv.square 0.9 Weight function 0.8 0.7 0.6 0.5 0 0.2 0.4 0.6 0.8 1 Edge probability Figure 2: Weight functions. Table 2 Dataset synopsis (from [5]). Property 1821 graph US2020 graph Number of vertices 132.317 147.881 Number of edges 2.225.177 2.447.224 Density / Log-density 16.8170 / 1.2393 16.5486 / 1.2357 Completeness / Log-completeness 2.54𝑒−4 / 0.6196 2.38𝑒−4 / 0.6173 Number of triangles 446.513 489.773 Number of squares 215.387 218.633 Number of cliques of size four 102.044 125.806 Graph diameter 10 11 Percentage of vertices reachable at diameter-1 95.33% 98.17% Percentage of vertices reachable at diameter-2 93.26% 96.44% Percentage of vertices reachable at diameter-3 89.11% 91.22% Percentage of vertices reachable at diameter-4 84.73% 87.47% Number of favorites 36.994.815 42.114.509 Number of tweets 17.465.844 22.773.674 Table 3 Parameters of the experiments. Parameter Meaning Value 𝑤𝑓 Edge follow weight 0.5 𝑤𝑟 Edge retweet weight 0.25 𝑤𝑐 Edge hashtag emotional coherence 0.25 𝜂0 State vector equality threshold 0.05 5 Georgios Drakopoulos et al. CEUR Workshop Proceedings 1–7 Table 4 Normalized number of iterations as a function of the hyperparameter 𝛽0 . Hyperparameter Graph inv.linear inv.linear+beh inv.square inv.square+beh 0.5 1821 1.49 1.26 1.37 1.38 0.7 1.45 1.24 1.32 1.23 0.8 1.39 1.16 1.27 1.15 0.9 1.33 1.08 1.22 1.07 1 1.28 1 1.19 1 0.5 US2020 1.41 1.17 1.33 1.13 0.7 1.42 1.14 1.29 1.11 0.8 1.37 1.09 1.25 1.08 0.9 1.29 1.05 1.21 1.03 1 1.24 1 1.18 1 Table 5 Emotional distributions (pos/neu/neg) computed with the best value of 𝛽0 . Graph init i.linear i.linear+beh i.square i.square+beh 1821 0.64/0.24/0.12 0.68/0.14/0.18 0.73/0.15/0.12 0.69/0.15/0.16 0.74/0.14/0.12 US2020 0.28/0.23/0.49 0.21/0.22/0.57 0.17/0.16/0.67 0.22/0.22/0.56 0.16/0.16/0.68 GNN is run with the latter removed from the initial local case. There it can be seen that the US2020 yields for ground truth vectors at the vertices. both weight choices slightly different results from the From table 4 it follows immediately that the inclusion initial distribution when the emotional factor is excluded of the behavioral factor in (4) leads to quicker conver- but considerably different ones when they are included. gence of the proposed GNN architecture. This can be Thus, it is a graph with a heavy emotional charge. On the attributed to the following reasons: other hand, the 1821 graph tends to yield similar results in every case, signifying thus greater coherency. • Information enrichment: From an algorithmic perspective, the behavioral factor adds an inde- pendent dimension to the profile of each vertex. 5. Conclusions And Future Work Hence, the new vertex profile space can differenti- ate adequately between dissimilar vertices while This conference paper focuses on a graph neural net- maintaining close enough similar ones. work architecture for discovering community structure • Numerical variation: The above is enhanced in large Twitter graphs. In this approach said structure is by having more diversified edge weights. Besides formed using a Twitter account behavioral model which the additional factor, the behavioral term is also results from fusing structural and functional attributes signed. In turn this expands the range of weights, with emotional ones. The proposed model can be natu- increasing thus the possible number of values. rally extended to include additional features from these categories or even ones belonging to different categories The above factors suggest that variability in the weight as long as they can be expressed in a numerical scale space as well as in the vertex profile increase the flexibil- where normalization does not influence semantics. In ity of the update mechanism of (11). This is in accordance our experiments the inclusion of behavioral attributes with the standard pattern recognition maxim stating that leads consistently to quicker GNN convergence. mapping data to a space of higher dimensionality facil- This work can be extended in a number of ways. First, itates their clustering. On the other hand, the curse of multiple weight functions can map each edge to a weight dimensionality imposes a limit on how big this new space vector and hence to a multidimensional weight space can get. As both spaces used in this work however are where each dimension has its own semantics. Then the low dimensional, this does not constitute a problem. fundamental parameters of candidate distributions de- Regarding the total sentiment, in table 5 is shown the scribing this space can be derived through signal esti- average emotional distribution before and after the GNN mation techniques. Second, alternative behavioral mod- execution in each case using the value of hyperparame- els depending only on local properties or on local esti- ter 𝛽0 which leads to the quickest convergence in each mates of global ones should be developed as this would be 6 Georgios Drakopoulos et al. CEUR Workshop Proceedings 1–7 most appealing for a distributed implementation. Third, [10] G. Drakopoulos, E. Kafeza, P. Mylonas, L. Iliadis, models for computing or estimating the edge existence Transform-based graph topology similarity met- probability which reflect the underlying graph semantics rics, NCAA 33 (2021) 16363–16375. doi:10.1007/ should be research objectives. Finally, given the evolv- s00521-021-06235-9. ing nature of online social networks, an architecture for [11] G. Drakopoulos, P. Mylonas, Evaluating graph re- non-stationary graphs should also be developed. In this silience with tensor stack networks: A keras im- case certain transient global graph states can be used to plementation, NCAA 32 (2020) 4161–4176. doi:10. obtain intermediate status results regarding clustering, 1007/s00521-020-04790-1. flow, degree distribution, or any other global property. [12] M. T. Schaub, S. Segarra, Flow smoothing and de- These transient states can well serve as starting points noising: Graph signal processing in the edge-space, for new partitioning techniques. in: GlobalSIP, IEEE, 2018, pp. 735–739. [13] A. Gavili, X.-P. Zhang, On the shift operator, graph frequency, and optimal filtering in graph signal pro- Acknowledgments cessing, IEEE Transactions on Signal Processing 65 (2017) 6303–6318. This conference paper was supported by the Research [14] F. Hua, R. Nassif, C. Richard, H. Wang, A. H. Sayed, Incentive Fund (RIF) Grant R18087 provided by Zayed A preconditioned graph diffusion LMS for adaptive University, UAE. graph signal processing, in: EUSIPCO, IEEE, 2018, pp. 111–115. References [15] S. Kontopoulos, G. Drakopoulos, A space efficient scheme for graph representation, in: ICTAI, IEEE, [1] C. Zhang, D. Song, C. Huang, A. Swami, N. V. 2014, pp. 299–303. doi:10.1109/ICTAI.2014.52. Chawla, Heterogeneous graph neural network, in: [16] T. M. Cihon, M. A. Mattaini, Emerging cultural ICDM, 2019, pp. 793–803. and behavioral systems science, Perspectives on [2] C. Yu, Y. Liu, C. Gao, C. Shen, N. Sang, Represen- behavior science 42 (2019) 699–711. tative graph neural network, in: ECCV, Springer, [17] I. Markovsky, F. Dörfler, Behavioral systems the- 2020, pp. 379–396. ory in data-driven analysis, signal processing, and [3] J. Kim, T. Kim, S. Kim, C. D. Yoo, Edge-labeling control, Annual Reviews in Control (2021). graph neural network for few-shot learning, in: [18] G. Drakopoulos, E. Kafeza, P. Mylonas, CVPR, 2019, pp. 11–20. H. Al Katheeri, Building trusted startup teams from [4] X. Fu, J. Zhang, Z. Meng, I. King, Magnn: Meta- LinkedIn attributes: A higher order probabilistic path aggregated graph neural network for hetero- analysis, in: ICTAI, IEEE, 2020, pp. 867–874. geneous graph embedding, in: WWW, 2020, pp. doi:10.1109/ICTAI50040.2020.00136. 2331–2341. [19] G. Drakopoulos, I. Giannoukou, P. Mylonas, [5] G. Drakopoulos, I. Giannoukou, P. Mylonas, S. Sioutas, On tensor distances for self organiz- S. Sioutas, A graph neural network for assessing ing maps: Clustering cognitive tasks, in: DEXA, the affective coherence of Twitter graphs, in: IEEE volume 12392 of Lecture Notes in Computer Sci- Big Data, IEEE, 2020, pp. 3618–3627. doi:10.1109/ ence, Springer, 2020, pp. 195–210. doi:10.1007/ BigData50022.2020.9378492. 978-3-030-59051-2\_13. [6] W. Fan, Y. Ma, Q. Li, Y. He, E. Zhao, J. Tang, D. Yin, [20] S. Kim, K. Song, B. Lockee, J. Burton, What is gamifi- Graph neural networks for social recommendation, cation in learning and education?, in: Gamification in: The WWW conference, 2019, pp. 417–426. in learning and education, Springer, 2018, pp. 25– [7] A. Ortega, P. Frossard, J. Kovačević, J. M. Moura, 38. P. Vandergheynst, Graph signal processing: [21] N. Wilkinson, M. Klaes, An introduction to behav- Overview, challenges, and applications, Proceed- ioral economics, Macmillan International Higher ings of the IEEE 106 (2018) 808–828. Education, 2017. [8] G. Mateos, S. Segarra, A. G. Marques, A. Ribeiro, Connecting the dots: Identifying network structure via graph signal processing, IEEE Signal Processing Magazine 36 (2019) 16–43. [9] M. Cheung, J. Shi, O. Wright, L. Y. Jiang, X. Liu, J. M. Moura, Graph signal processing and deep learning: Convolution, pooling, and topology, IEEE Signal Processing Magazine 37 (2020) 139–149. 7