FRAMING COWORKING SPACES MARKETING STRATEGIES VIA SOCIAL MEDIA INDICES

: In this paper an investigation of social media marketing techniques of Coworking spaces’ type of business is performed, using datasets acquired using social media monitoring tools. Mediatoolkit has been used to scrap data deriving from the activity of the WeWork Instagram and Twitter accounts which were collected on a 24/7 basis from varying locations and in multiple languages in a fifteen-day time window. Indices related to sentiment, reach, influence, number of followers, retweets, likes, comments, and view scores formed the datasets that were examined by applying multiple correspondence analysis as well as the hierarchical clustering method. The aim of this paper was to explore the inherent properties of the multiple indices describing the general realm of social media marketing tools, and more specifically aspires to provide digital marketers with an alternative perspective of social media marketing strategies related to the emerging coworking spaces type of business. The authors identified three classes/segments of posts, whereas post polarity tends to relate to geographic location, regardless of the social media channel used for posting.


Introduction
Data analytics can be applied in a variety of cognitive fields with marketing being traditionally among them. Advances of technology and digital media have led to the 3 beyond traditional marketing capabilities (Alves et al., 2016). Social media marketing is, therefore, a prerequisite for building online brand communities, driving leads and sales, improving business performance and sustaining their competitive position on the market.
Moreover, globalization and digital technologies have substantially transformed the world of work (Ivaldi, Pais, and Scaratti, 2018). Working has became possible at any place and time, and the stable employment paradigm is beginning to collapse (Spinuzzi, Bodrožić, Scaratti, and Ivaldi, 2019). Subsequently, remote work has spread and the number of self-employed workers, as well as freelancers, has increased (Weijs-Perrée, van de Koevering, Appel-Meulenbroek, and Arentze, 2019, pp. 534--548). The global workforce started to change, and economies are becoming more flexible, agile, entrepreneurial and innovative. Such market trends affected demand for office space as the need arose for a flexible working environment, other than traditional or private setting (Van de Koevering, 2017). Under these circumstances, coworking spaces provided a solution to the situation, and the proliferation of coworking spaces emerged at the onset of the structural changes in the labour market (Gandini, 2015;Merkel, 2015).
Coworking represents a new way of working expressed as individual work in a shared environment (Constantinescu and Devisch, 2018). Coworking space can be therefore understood as a shared workspace where professionals from different business backgrounds work alongside one another (Merkel, 2015;Weijs-Perrée et al., 2019). Thus it stimulates networking, collaboration and the development of a professionally heterogeneous community (Kubátová, 2014). Since its emergence in 2005 in San Francisco, the coworking sector has undergone rapid growth (Foertsch, 2019;Kubátová, 2014;Leclercq-Vandelannoitte and Isaac, 2016). The number of coworking spaces has globally risen from 600 in 2010 to 18700 open spaces in 2018, with 1 650 000 workers using coworking space (Statista, 2018a(Statista, , 2018b. This rapidly growing phenomenon transformed the way people work and collaborate (Leclercq--Vandelannoitte and Isaac, 2016, pp. 3-9).
As pointed out above, data analytics can be applied in the field of social media marketing, exploiting the vast amount of data available in the numerous social media platforms towards shaping effective marketing techniques. Collaboration seems to be the common factor between social media applications and coworking spaces. This paper attempts to extract and analyse social media marketing data originating from the specific field of coworking spaces, and more specifically from the WeWork company's social media presence. The extracted datasets were analysed by applying multiple correspondence analysis and hierarchical clustering method in an effort to investigate whether these data analysis methods can support planning digital marketing strategies.

Literature review
Data mining and the methods of Big Data studies have attracted a great deal of attention in the information industry in recent years, due to the wide availability of huge amounts of data in the social web and the urgent need for turning such data into useful knowledge (Gürsoy, Bulut, and Yigit, 2017). The potential of social media Big Data is colossal in marketing and with data being generated and collected in real--time, round the clock, seven days a week, the marketing industry is now able to see what people are buying, following or communicating about (Barutcu, 2017). The visualization of this kind of mined data has opened up new ways in evolving marketing decision models (Treboux et al., 2016). While social network analysis (Kefi, Indra, and Abdessalem, 2016) and content analysis can be one approach, the introduction of indexes that describe the social web's parameters have opened up new techniques of statistical research for a variety of fields. For example, Thoma et al. used a social media index containing four metrics to measure statistically the impact of emergency medicine and critical care websites (Thoma et al., 2015). Arora et al. introduced a mathematical model for social media index valuation involving technological, social, economic, and ethical dimensions in their attempt to create benchmarks signifying a company's share and commitment to social media (Arora, Arora, and Palvia, 2014). Zhou et al. also applied social media analysis to reveal collective behaviour based on data and indices retrieved from Twitter and Sina Weibo, the most popular microblogging services all over the world and in China, respectively (Zhou, Qian, and Ma, 2012). Bucko et al. used statistical methods to identify the factors that affect consumer purchasing behaviour by applying principal components analysis and factor analysis to their decision criteria (Bucko, Kakalejčík, and Ferencová, 2018). Advances in machine learning techniques raised the popularity of Sentiment Analysis which started to be used as a separate mature index in Social Media investigations generally (Iglesias and Moreno, 2019), as well as in quantitative and qualitative marketing research methods enriching the datasets under investigation (Rambocas and Gama, 2013). Markic et al. showed that sentiment analysis can be used to create marketing strategies and improve customer relations and customer service (Markic et al., 2016). Indices like reach, passion and strength have also been used in previous statistical analysis in marketing issues (Vagianos and Koutsoupias, 2017). In this kind of research, multiple correspondence analysis (MCA) is a common technique for identifying groupings, or market segmentation, in which groups of people with similar characteristics are considered together (Diana and Pronello, 2010). Furthermore, in recent decades cluster analysis has become a common tool for the marketing researchers, helping to create groupings of variables of the issue under investigation (Punj and Stewart, 1983).
In this paper, indices related to social media activity generated by the WeWork company that have to do with sentiment, reach, influence, number of followers, retweets, likes, comments and view scores, formed the datasets that have been Framing coworking spaces marketing strategies via social media indices 5 investigated by applying multiple correspondence analysis (MCA) as well as hierarchical clustering (HC). This approach has been used extensively in the research on marketing of services and products in numerous cases, such as the combined use of MCA and HC was employed by dos Santos et al. (2020) for the exploration of the insurance market. Similarly, Wen and Chen (2010) explored the competitive positions of international air travellers, Brida et al. (2014) studied sociodemographic and travel-related characteristics of cruises, whilst Diana and Pronello (2010) defined a set of different customer profiles regarding travellers' social status and characteristics and efficiency of transportation networks. Bejaei, Cliff, and Singh (2020) and Nicolosi, Fava, and Marcianò (2018) examined purchasing habits and preferences for fruit and fishery, respectively.
The procedure aimed to highlight social media Big Data from a new perspective which can help companies from this sector to effectively plan and adjust their marketing strategy.

Dataset and variables
In this paper, the dataset for analysis was retrieved using Mediatoolkit, an online media monitoring software that enables marketers and organizations to be notified about every mention of their company or product. It monitors over 10 million websites and 100 million social media profiles, helping brands to manage their online reputation. Whenever Mediatoolkit recognizes a mention, it immediately sends out a real-time alert that can be found in the email, online on the website or directly in the phone application (Mediatoolkit, 2019). This feature can be very helpful in the case of a negative mention because it can be put right without delay, not disrupting a brand's image to such an extent as if it had been left without notice for a longer period (FinancesOnline, 2019). The trial version of the platform was used in a 14-day window (6 th -20 th January 2019), in order to extract spreadsheets containing datasets as the results of mention queries. This version allowed for carrying out up to 100 queries, showing up to 50 000 mentions, tracking keywords or phrases and monitoring or tracking on specific channels, such as Facebook, Twitter, Instagram, YouTube and general Websites.
It also allowed for choosing only mentions of specific social media platforms with a specific language, location, authors or sentiment.
Mediatoolkit was used in order to investigate the social media activity performed by the WeWork Company. WeWork is a private American company that provides shared workspaces for entrepreneurs, start-ups and freelancers, as well as for small and big businesses. In just nine years of its existence, WeWork has created a global network of coworking spaces present in 660 locations, 115 cities and 27 countries. It manages 45m square feet of office space globally where more than 400 000 members can enjoy 100000 annual community events (Sullivan, 2018;WeWork, 2019). Therefore, in its early stages it could be viewed as a suitable example of a coworking spaces company with a proven ability to expand rapidly in a short period of time.
Eventually, five queries were executed over Instagram, Twitter and the web involving the following keywords or hashtags: #dogsofwework, ##dowhatyoulove + #wework, #wearewework, #wework and WeWork. The queries provided results relevant to mentions over time (Table 2), total impressions, positive-negative sentiment ratio and sentiment over time.
Specifically, they provided datasets involving the following variables: auto_ sentiment, reach, influence score, followers, like count, comment count, favorite count, view_count and retweet count. There were also metadata indicating details like the title length, the country of origin, the spoken language and whether there is an image or not in the posts. These variables were then classified into two main categories (Table 2): quantitative and qualitative. The final selected data set was formed using MS Excel and R language base functions and is a complete table of 3967 rows and nine columns, since posts with no defined location, language and sentiment were excluded. The dataset with the variables described above provided the input for applying the MCA as well as the HC method described in the next paragraphs.

Methodology and results
Recent surveys are increasingly collecting nonmetric data via categorical variables like opinion posts, difficult to analyse with the most frequently applied tools. Thus, as mentioned before, the methodology chosen for this data set exploration was a combination of multiple correspondence analysis in tandem with hierarchical clustering (Ward's metric) using the R package FactoMineR implementation (Le et al.). As with factor fnalysis and principal component analysis in parametric experiments, MCA reduces the dimensionality of results. Moschidis, Chatzipetrou, and Tsiotras (2018) provided detailed explanations of the combined use of MCA and HC. Further on, the mathematics on both methods is purposefully left out of this article. The statistical aspects of MCA can be found in the works of Benzecri (1992) and Clausen (1998), while the technicalities of HC can be found in Mirkin (2012).
The application of MCA assumes all examined variables are categorized, thus the TIT, MEN, REA and FOL values were initially transformed to categories based on the Q1, Q2-Q3, Q4 rule (Table 3). On the basis of the MCA output, the amount of inertia (total information revealed) in the examined data along the first axis is 19% and along the second axis is 13.6% (Figure 1). This is the first view of the examined data on a multi-variate basis, revealing the customer opinion main trends.
The    The Euclidean distance was calculated between individuals, and Ward's criterion was applied as the clustering method to minimize the within-cluster variance (Husson, Josse, and Pages, 2010). Furthermore, a chi-squared test (χ2 test) was performed automatically via the implementation of MCA in the FcrtoMineR package to identify the variables that characterized the clusters. The chi-squared value is equal to (df). The positioning of customer posts can also be depicted according to the clustering of post groups using the corresponding R tools, namely functions included in the factoextra package (Kassambara and Mundt, 2017). Each cluster is characterized by a distinct group of categories revealing inherent customer properties.

Results and discussion
The resulting clusters based on HC output have the following characteristics: a) Cluster 1 (n = 1454): represents neutral Instagram post mostly from Japan, Belgium, Italy Lithuania, China, India, Estonia, Turkey, Finland, Russia, Israel, Georgia, Slovakia, Portugal, Czechia, South Africa, and Hong Kong. b) Cluster 2 (n = 1988): corresponds to the most populated group, including posts on Instagram, YouTube and the Web. The countries present here are the US, the Netherlands, Australia, Saudi Arabia, Ethiopia, Ecuador, Armenia, Peru, Korea (the Republic of), Brazil, and Slovenia. All with mainly positive posts. c) Cluster 3 (n = 525): the smallest group includes mainly negative Instagram posts from Argentina, Sweden, Mexico, Denmark, Colombia, Jamaica, Ghana, Venezuela, Uruguay, and Somalia.
Thus, the employed multi-variable techniques clearly depict formations and clusters in the examined phenomenon, providing a full view. As shown earlier, in terms of service design and/or brand positioning/repositioning, the MCA axes ( Figure 1) may initially provide key input to services marketers and brand managers. Additionally, HC may be used to assess which groups are thought to be closer to each other (Figure 3), in addition to the perceptual axes created by MCA. The general process flow of our methodological approach is depicted in Figure 4.

Conclusion
The combined use of MCA and HC defined three types of posts and revealed that, regardless of the social media channel used for posting, post polarity is mainly related to location. Each of these three groups of posts has distinct geographical characteristics, and online retailers may use this knowledge strategically to target their audiences more efficiently. Therefore, this study showed that the combination of these methods may serve as explorative preliminary research to investigate the interrelationship between posts and overall polarity on a per-country basis.
It must be noted that the selected time window of January 2019 provided pre--pandemic data that correspond to a time of normal prevailing circumstances. Recently, there has been a lot of debate about the impact of the pandemic on directions of various business models, and coworking spaces could be definitely among them. More recent time windows could provide biased results due to this fact. The validity of the results is further enhanced by the fact that the selected time window belongs to the period before the company's spectacular financial collapse in September 2019, which was due to a variety of reasons. It is more than evident that social media data retrieved over the months of the company's downfall would lead to misguided conclusions.
What exactly is the nature and depth of commitment reflected by the local coworking spaces industries at country level can be recommended for future research. Researchers adopting the proposed method (or other similar methods) should certainly take into account the limitations posed by the VPN services or chatbots over the validity of the sentiment analysis results.