Nowadays, more and more people are using online social media to express their thoughts and opinions on a variety of topics that interest or concern them. Through social networking platforms, people have the ability to communicate directly and share knowledge with people all around the world. Twitter is one of the most popular social media, used by millions of users daily. In particular, people use it to express their opinion directly and freely on whatever concerns them, thereby generating a large amount of data. The abundance of this information and its multifaceted importance, emphasizes how important is to find ways of collecting and analyzing such data in order to extract valuable knowledge. Such data, are a valuable source of information whose extraction can help individuals or even businesses in the decision-making process. The present research focuses on the study of user communication about a brand on Twitter, and in particular on exploiting user feelings about this brand efectively. In more detail, this work promotes the eficient modeling and management of the business-consumer relationship by studying the interactions of users who are discussing a specific brand name. The purpose of this research work is to provide an eficient tool that will enable businesses to use technological and automated tools in order to efectively manage the emotional state of consumers in relation to their brand. Consumer feedback and expressed emotions may be utilized by companies for making decisions regarding marketing research, competitive business intelligence and online reputation management.
sumers’ thoughts and opinions about a brand name afect their behavior and consequently their brand perception.
It is widely known that the evolution of the Web 2.0 According to [2, 3] the consumers trust others’ opinions, has led into a new era, where the social networks have thus the electronic word-of-mouth (eWOM) has a magained a crucial role in people’s lives. More specific, in jor role on how users perceive a brand name. The users the recent years, Twitter has been one of the most pop- are no longer receivers of information but they have beular networks that has been broadly utilized in a wide come transmitters. Thus, there is a considerable need plethora of research studies that are trying to extract and from businesses to assess their relationships with the cusanalyze users’ activity with the intention of finding valu- tomers by developing eficient quality and quantitative able trends or patterns [1]. Such trends or patterns have metrics on the social networks [4], in order to be able to been proved to be valuable information for businesses. create and maintain their customer relationships.
On the one hand, users nowadays are informed regard- Although there are several tools that look into the probing businesses’ products and services through the social lem of social media networks and their interconnections, networks and they tend to interact with other users to in this work we focus on the extraction of emotional inexchange information, opinions and discuss about prod- formation in order to relate it to the loyalty of the users ucts. On the other hand, businesses through the social to a specific brand. So, the scope of this research work is networks tend to promote and advertise their products or to provide a valuable tool for the eficient modelling and services to a wide range of consumers. Businesses have management of the business-consumer relationships by also gained the capability to observe the impact of their analysing the interactions of users who are discussing a products through consumers’ opinions as well. specific brand name.
Therefore, the study of the interactions of consumers The rest of this paper is organized as follows. Section and the relationships they develop with brand names on 2 presents the related work regarding multilayer graphs, social networks, has become vital for businesses. Con- graph signal processing (GSP), and emotional analysis over online social networks. The main concepts of our CIKM’21: 30th ACM International Conference on Information and system architecture are covered in Section 3. In Section Knowledge Management, November 01–05, 2021, Virtual Event, QLD, 4, we present our experimental results on the perforAustralia mance of our proposed methodology based on a variety $ robolas@ceid.upatras.gr (G. Rompolas); of metrics. Finally, in Section 5, we conclude the paper kkaravoul©ia20@21 cCeopiydri.guhtpfaorttrhaissp.agprer(bKyi.ts Kauathroarsv.Uosuelpiearm)itted under Creative by outlining our findings and discussing on the future CPWrEooUrckReshdoinpgs IhStpN:/c1e6u1r3-w-0s.o7r3g CCoEmmUoRns LWiceonsrekAstthribouptionP4r.0oIncteerenadtiionnagl s(CC(CBYE4U.0)R.-WS.org) work.
partitioning of our network graph into two smaller subgraphs based on the type of relationships between the Multilayer or multiplex graphs constitute a graph class users. where vertices can be connected with multiple edges Then, we are analysing both of these graphs indepenas long as they have distinct labels [5]. Operations on dently and we are applying a scalable community desuch graphs include embedding [6] and core decompo- tection algorithm to further divide them based on the sition [7]. Moreover, clustering in such graphs can be multilayer network structure. And finally, we are doing spectral through convex aggregation [8], local graph con- a community emotion analysis on each cluster in order volution [9], or through the respective Laplacian [10]. to gain further insights of the attributes and emotions Applications include brain circuit study [11], unmixing that characterize each one of the extracted communities. of hyperspectral images through adaptive non-negative Each one of the aforementioned steps, are being further matrix factorization [12], and race car trajectory plan- analysed in the following sections. ning for transportation networks [13]. Multilayer graphs can also model social networks facilitating the solution of problems such as link prediction [14], higher order vertex centrality [15], account behavior prediction [16], suspicious activity detection [17], and stable community detection [18].
GSP is an emerging field [ 19][20] with numerous applications including graph partitioning with methods such as spectral clustering as in [21], submodular computation [22], graph dimensionality reduction [23], higher order iterative methods [24], multi-view [25], and vertex Figure 1: Architecture Overview search [26]. Additionally, GSP covers the scenario where a neural network architecture such as marginalized graph autoencoders [27], tensor stack networks (TSNs) [28], or 3.1. Data crawling graph neural networks (GNNs) [29] is applied to a graph for processing purposes such as clustering or community We collect data by crawling the Twitter. We are interested structure discovery [30]. in customers-business relationships, therefore we select
Social media in general and Twitter in particular of- those tweets that include either the brand name or the fer numerous opportunities for observing the emotional hashtag of the the brand name. Thus we extract users that evolution of human interaction [31]. The recent bibliog- are potential customers since they possess basic brand raphy abounds with approaches ranging from employing awareness. information difusion patterns in conjunction with textual information [32] and transformers applied on tweets 3.2. Graph Construction for deep text mining [33] to neural network architectures [34] and ensemble classification [ 35]. Applications A Twitter graph has nodes that represent the users and include among others cryptocurrency price prediction edges that usually represent the follow relationships bebased on emotional attributes [36], the behavioral dynam- tween them. However, we are interested to represent ics of product customers on Twitter [4], the impact of the users’ interactions in a more detailed manner than Twitter collective sentiment on the price of energy stocks representing only the follow relationships. Therefore, we [37], and well being [38]. Finally, an extensive review of consider three types of interactions: current emotional analysis techniques for Twitter is [39].
In Figure 1, an overview of our proposed model architecture is depicted. Initially, in order to apply our model it is necessary to crawl the corresponding data; in our case we need to collect data regarding a brand name. Then we proceed with the graphs construction, where we are building the diferent graph layers that are needed to construct a multilayer graph. Since the multilayer graph construction has been completed, we proceed with the
2. Retweets (RT ) 3. Replies (RP)
For each type of interactions we construct a directed graph (, ), where is the vertex set, which contains all the network users, is the edge set and ∈ { , , }. We consider a directed edge from the node to if the user re-tweets, mentions or replies to the user .
According to the Section 3.2, we have constructed three independent network graphs, one for each user interaction. Therefore, since each network has its own structure and its own characteristics that define it, an appropriate mechanism to represent this information are the multilayer graphs [40, 41]. A multilayer graph can capture the several interactions that can exists among the nodes of each layer, without losing any information.
users we maintain only on the common ones among all the layers. We combine the three layers into a multiplex network. Multiplex networks have the same entities in every layer but diferent connections between them. Thus, we transform the three diferent interaction graphs as defined in Section 3.2 into a multiplex multilayer Twitter graph (, , ), where is the vertex set, is the edge set and ∈ { , , }. 3.4. Multilayer network partitioning community. We also defined as input to the algorithm the number of communities = 2. Hence, in our case, after the execution of the algorithm two distributions are being produced for every node, one for each community. Although the algorithm supports the nodes to be assigned to more than one community, we use hard node assignments in our approach, where a node belongs to a single community. Therefore, we apply the MULTITENSOR algorithm to each one of the two networks, in order to extract two communities from each case.
In order to gain further insights from our extracted communities and translate them into meaningful information, we use the state-of-the-art NRC Emotion Lexicon[43, 44] that categorizes the emotions in the following emotions {anger, anticipation, disgust, fear, joy, sadness, surprise, trust} and sentiments {positive, negative}. Based on the above afect categories we quantify the emotions of the tweets at a finer granularity level.
3.8.0 as the programming language. The python library Tweepy1 was used as a wrapper to access the Twitter API2 to retrieve data from Twitter. The MULTITENSOR3 library was used as well, with regards to the multilayer tensor factorization for community detection.
Upon the construction of the multilayer networks, we proceed with its partitioning into two smaller multilayer For our implementation, we selected the Adidas sub-networks, based on the type of the users’ commu- sportswear brand name. Thus we collected Tweets that nication relationships. More specific, we separate our included the word Adidas or the hashtag #Adidas for initial network into a sub-network (, , ) that the time interval from 09-07-2019 till 29-07-2019. We represents the interactions between users who communi- maintained only the tweets that were in the English lancate about the brand name, and into another sub-network guage. Moreover, we filtered out users mentions, the RT (, , ) that represents the user interactions with abbreviation symbol that denotes a retweet, hyperlinks, the oficial brand name accounts on Twitter. numbers, symbols and punctuation as well.
Subsequently, we apply on the multilayer graph the The User-User network is composed of = 3618 users MULTITENSOR[42] community detection algorithm. and 12620 interactions. If we define the density of the The algorithm takes as input the adjacency matrix of 1ℎ : //ℎ.// a multilayer network and returns for every node of the 2ℎ : //../ network the probability distributions of the node for each 3ℎ : //ℎ.// (a) Emotional comparison between the two networks (b) Tree-map of the overall emotions interactions. There were two communities created where the first one had 357 nodes and the other 1157 accordingly. Table 4.4 shows respectively some structural information of the network. We can notice that although the User-Adidas network has much fewer nodes, it seems to be a more dense network than the User-User network.
Furthermore, the clustering quality was very good in this case as well, with = 0.98 and 1 = 0.017.
D Mentions (MT)
Replies (RP) Retweets (RT)
E = 1 ∑=︁1 |00||| (2) tcUialpatoiyonenrthonefeetcwaocmohrmkcsou,mnwimteieupsnreioxtcyter.aeTcdhteieodnFtihfgreuorememb3ootstihohonowsfsqthuoeaunrmtaiufi-gl-where || denotes the Euclidean norm. We also com- gregated results. Initially, we measured the mean values puted the 1 metric as follows: of all the users’ emotions mentioned in Section 3.6 over each one of the multilayer graphs. It is interesting to 1 = 21 ∑=︁1 ||0 − ||1 (3) anMnoodtirceleeosvtsheiarn,ttwearletahcaotliusoognhsdtiitshteiisnAagdumiidsoahrseedneemtthwoetoioorkvneahrlaaislnllfeeosmrsmonatoitoidonenss.
The value of = 0.99 which is value near 1, while that users express regarding the brand name. In our case, 1 = 0, 020. Therefore the quality of the clustering is the anticipation, the joy and the trust emotions and the very good. sentiments are the most dominant ones, as they have the largest mean values, Figure 3. 4.4. The User-Adidas interaction network We also measured the mean values of all the emotions over each one of the extracted communities for each mulThe User-Adidas network consists of = 1514 users, tilayer network. In the case of the User-User network we including 16 accounts of the Adidas Company and 4463 (a) User-User Network (b) User-Adidas Network observe that the emotional diferences between the clusters are not significant, as we can notice from the Figure 4(a) that for each emotion the mean values are almost equally. Applying the same approach for the User-Adidas network we observe that the emotions of the two clusters are significantly diferent, Figure 4(b). Thus, in this case the community detection algorithm has achieved a better separation of the users, regarding their emotions.
In order to further validate our results, we used the T-test [45] to compare the mean values of the clusters with respect to emotions. In the case of the User-User network the value of the T-test was 0.111, which indicates the similarity of the extracted communities. While in the case of the User-Adidas network the value of the T-test was 1.009, that proves that the two clusters are quite diferent with respect to emotions.
Moreover, a visual representation of the two graphs can verify our findings. In Figure 5, the two diferent multilayer networks are depicted, where each user is labelled to its respective community. The Figure 5(a) refers to the User-User network where it is clear that the network is separated to two clusters of diferent emotions, while the Figure 5(b) refers to the User-Adidas one, where the clusters do not show significant diferences.
for the Adidas brand name. Two diferent networks were examined, that represent diferent types of interactions among the users. The first one represents the users’ interactions and the second one the interactions with the Adidas accounts. We used multilayer graphs to represent these interactions. We then applied a community detection algorithm to determine user communities. Finally, we examined the users’ emotions so that we can diferentiate each network in two sub-networks based on the [9] C. Wang, B. Samari, K. Siddiqi, Local spectral emotions. graph convolution for point set feature learning,
Through our methodology, we reached to the follow- in: ECCV, 2018, pp. 52–66. ing conclusions. The interactions among the users does [10] P. Mercado, A. Gautier, F. Tudisco, M. Hein, The not provide us with as much valuable information as the power mean Laplacian for multilayer graph clusinteractions between the users and the brand name ac- tering, in: International Conference on Artificial counts. The community detection algorithm achieved a Intelligence and Statistics, PMLR, 2018, pp. 1828– better emotion separation based on the latter interactions. 1838.
Furthermore, among all the emotions that were analysed [11] K. Mandke, J. Meier, M. J. Brookes, R. D. O’Dea, through this work, we were able to distinguish some par- P. Van Mieghem, C. J. Stam, A. Hillebrand, ticular ones as the most notable, such as the anticipation, P. Tewarie, Comparing multilayer brain networks the joy and the trust emotions and the sentiments. between groups: Introducing graph metrics and rec
As a possible future direction, it would be interesting ommendations, NeuroImage 166 (2018) 371–384. to apply our methodology on the study of how the users’ [12] L. Tong, J. Zhou, B. Qian, J. Yu, C. Xiao, Adapinteractions and emotions change over time. Moreover, tive graph regularized multilayer nonnegative maan additional potential approach could be to focus on trix factorization for hyperspectral unmixing, IEEE the development of prediction models that will take into Journal of Selected Topics in Applied Earth Obserconsideration the users’ behavior and emotions in order vations and Remote Sensing 13 (2020) 434–447. to predict the future relationships between consumers [13] T. Stahl, A. Wischnewski, J. Betz, M. Lienkamp, and businesses. Multilayer graph-based trajectory planning for race vehicles in dynamic scenarios, in: ITSC, IEEE, 2019, pp. 3149–3154.
References [14] H. Mandal, M. Mirchev, S. Gramatikov, I. Mishkovski, Multilayer link prediction in [1] A. H. Zadeh, R. Sharda, Modeling brand post popu- online social networks, in: TELFOR, IEEE, 2018, larity dynamics in online social networks, Decision pp. 1–4.
Support Systems 65 (2014) 59–68. [15] G. Drakopoulos, Tensor fusion of social structural [2] J. Berger, R. Iyengar, Communication channels and functional analytics over Neo4j, in: IISA, IEEE, and word of mouth: How the medium shapes the message, Journal of consumer research 40 (2013) [16] 2D0.1P6e.rdnoai,:1R0. .In1t1e0rd9o/nIaItSoA,.A2. 0T1a6g.ar7e7ll8i,53Id6e5n.tifying 567–579. users with alternate behaviors of lurking and active [3] C. Kudeshia, A. Kumar, Social eWOM: Does it af- participation in multilayer social networks, IEEE fect the brand attitude and purchase intention of Transactions on Computational Social Systems 5 brands?, Management Research Review (2017). (2017) 46–63. [4] E. Kafeza, C. Makris, G. Rompolas, F. Al-Obeidat, [17] P. Bindu, P. S. Thilagam, D. Ahuja, Discovering Behavioral and migration analysis of the dynamic suspicious behavior in multilayer social networks, customer relationships on Twitter, Information Computers in Human Behavior 73 (2017) 568–582.
Systems Frontiers 23 (2021) 1303–1316. [18] W. Rebhi, N. B. Yahia, N. B. B. Saoud, Discovering [5] F. McGee, M. Ghoniem, G. Melançon, B. Otjacques, stable communities in dynamic multilayer social B. Pinaud, The state of the art in multilayer net- networks, in: WETICE, IEEE, 2018, pp. 142–147. work visualization, in: Computer Graphics Forum, [19] G. Mateos, S. Segarra, A. G. Marques, A. Ribeiro, volume 38, Wiley Online Library, 2019, pp. 125–149. Connecting the dots: Identifying network structure [6] W. Liu, P.-Y. Chen, S. Yeung, T. Suzumura, L. Chen, via graph signal processing, IEEE Signal Processing Principled multilayer network embedding, in: IEEE Magazine 36 (2019) 16–43.
International Conference on Data Mining Work- [20] A. Ortega, P. Frossard, J. Kovačević, J. M. Moura, shops ICDMW, IEEE, 2017, pp. 134–141. P. Vandergheynst, Graph signal processing: [7] B. Liu, F. Zhang, C. Zhang, W. Zhang, X. Lin, Core- Overview, challenges, and applications, Proceedcube: Core decomposition in multilayer graphs, in: ings of the IEEE 106 (2018) 808–828. International Conference on Web Information Sys- [21] J. Wen, Y. Xu, H. Liu, Incomplete multiview spectems Engineering, Springer, 2020, pp. 694–710. tral clustering with adaptive graph learning, IEEE [8] P.-Y. Chen, A. O. Hero, Multilayer spectral graph Transactions on cybernetics 50 (2018) 1418–1429. clustering via convex layer aggregation: Theory [22] P. Li, O. Milenkovic, Submodular hypergraphs: pand algorithms, IEEE Transactions on Signal and Laplacians, Cheeger inequalities and spectral clusInformation Processing over Networks 3 (2017) 553– tering, in: ICML, PMLR, 2018, pp. 3014–3023. 567. [23] X. Zhu, S. Zhang, Y. Li, J. Zhang, L. Yang, Y. Fang, Low-rank sparse subspace for spectral clustering, sentiment on renewable energy stocks, Energy IEEE Transactions on Knowledge and Data Engi- Economics 76 (2018) 153–169.
neering 31 (2018) 1532–1543. [38] R. A. Plunz, Y. Zhou, M. I. C. Vintimilla, K. Mcke[24] Y. Zhao, Y. Yuan, F. Nie, Q. Wang, Spectral cluster- own, T. Yu, L. Uguccioni, M. P. Sutto, Twitter sentiing based on iterative optimization for large-scale ment in New York City parks as measure of welland high-dimensional data, Neurocomputing 318 being, Landscape and urban planning 189 (2019) (2018) 227–235. 235–246. [25] Z. Kang, G. Shi, S. Huang, W. Chen, X. Pu, J. T. Zhou, [39] D. Zimbra, A. Abbasi, D. Zeng, H. Chen, The stateZ. Xu, Multi-graph fusion for multi-view spectral of-the-art in Twitter sentiment analysis: A review clustering, Knowledge-Based Systems 189 (2020). and benchmark evaluation, TMIS 9 (2018) 1–29. [26] G. Drakopoulos, E. Kafeza, P. Mylonas, [40] M. Kivelä, A. Arenas, M. Barthelemy, J. P. Gleeson, H. Al Katheeri, Building trusted startup teams from Y. Moreno, M. A. Porter, Multilayer networks, JourLinkedIn attributes: A higher order probabilistic nal of complex networks 2 (2014) 203–271. analysis, in: ICTAI, IEEE, 2020, pp. 867–874. [41] M. De Domenico, A. Solé-Ribalta, E. Cozzo, doi:10.1109/ICTAI50040.2020.00136. M. Kivelä, Y. Moreno, M. A. Porter, S. Gómez, A. Are[27] C. Wang, S. Pan, G. Long, X. Zhu, J. Jiang, Mgae: nas, Mathematical formulation of multilayer netMarginalized graph autoencoder for graph cluster- works, Physical Review X 3 (2013) 041022. ing, in: CIKM, 2017, pp. 889–898. [42] C. De Bacco, E. A. Power, D. B. Larremore, C. Moore, [28] G. Drakopoulos, E. Kafeza, P. Mylonas, L. Iliadis, Community detection, link prediction, and layer Transform-based graph topology similarity met- interdependence in multilayer networks, Physical rics, NCAA 33 (2021) 16363–16375. doi:10.1007/ Review E 95 (2017) 042317.
s00521-021-06235-9. [43] S. M. Mohammad, P. D. Turney, Crowdsourcing a [29] R. Chandra, T. Guan, S. Panuganti, T. Mittal, U. Bhat- word–emotion association lexicon, Computational tacharya, A. Bera, D. Manocha, Forecasting trajec- Intelligence 29 (2013) 436–465. tory and behavior of road-agents using spectral [44] S. M. Mohammad, P. D. Turney, Emotions evoked clustering in graph-lstms, IEEE Robotics and Au- by common words and phrases: Using mechanical tomation Letters 5 (2020) 4882–4890. Turk to create an emotion lexicon, in: Proceed[30] F. Scarselli, M. Gori, A. C. Tsoi, M. Hagenbuchner, ings of the NAACL HLT 2010 workshop on comG. Monfardini, The graph neural network model, putational approaches to analysis and generation IEEE Transactions on neural networks 20 (2008) of emotion in text, Association for Computational 61–80. Linguistics, 2010, pp. 26–34. [31] Z. Jianqiang, G. Xiaolin, Z. Xuejun, Deep convolu- [45] A. Kak, Evaluating information retrieval algorithms tion neural networks for Twitter sentiment analysis, with significance testing based on randomization IEEE Access 6 (2018) 23253–23260. and students paired t-test, Tutorial Presentation [32] L. Wang, J. Niu, S. Yu, Sentidif: Combining textual (2013).
information and sentiment difusion patterns for Twitter sentiment analysis, IEEE Transactions on Knowledge and Data Engineering 32 (2019) 2026– 2039. [33] U. Naseem, I. Razzak, K. Musial, M. Imran, Transformer based deep intelligent contextual embedding for Twitter sentiment analysis, Future Generation Computer Systems 113 (2020) 58–69. [34] A. S. M. Alharbi, E. de Doncker, Twitter sentiment analysis with a deep neural network: An enhanced approach using user behavioral information, Cognitive Systems Research 54 (2019) 50–61. [35] N. Saleena, An ensemble classification system for
Twitter sentiment analysis, Procedia computer science 132 (2018) 937–946. [36] O. Kraaijeveld, J. De Smedt, The predictive power of public Twitter sentiment for forecasting cryptocurrency prices, Journal of International Financial
Markets, Institutions and Money 65 (2020). [37] J. C. Reboredo, A. Ugolini, The impact of Twitter