Meetup.com is a global online platform which facilitates the organisation of meetups in di erent parts of the world. A meetup group typically focuses on one speci c topic of interest, such as sports, music, language, or technology. However, many users of this platform attend multiple meetups. On this basis, we can construct a co-membership network for a given location. This network encodes how pairs of meetups are connected to one another via common members. In this work we demonstrate that, by applying techniques from social network analysis to this type of representation, we can reveal the underlying meetup community structure, which is not immediately apparent from the platform's website. Speci cally, we map the landscape of Dublin's meetup communities, to explore the interests and activities of meetup.com users in the city.
Meetup.com is an online platform that helps people to plan, organise, and discover public or private events, referred to as meetups. Members of the platform can join speci c meetup groups, that are usually based around a speci c topic or activity, within which meetup events are organised. Meetup.com hosts groups that focus on diverse topics including sports, food, language, technology, business, philosophy, and dancing. The meetup.com platform is used worldwide and at the time of writing hosted more than 300k meetup groups and 39m users. On average over 3m people attend a meetup event each month.
A complex network structure underlies the meetup.com platform, consisting
of connections between users, meetup groups, and meetup events. By applying
popular techniques from the eld of social network analysis, such as centrality
analysis [
In this paper, we describe an analysis of the meetup.com network in Dublin,
Ireland. Rather than focusing on individual users of the platform, we instead
form a network representation at the meetup level, where a connection between
two meetups exists if the two groups share members in common. We focus on two
key research questions: 1) do distinct thematically-coherent communities exist
within Dublin's Meetup ecosphere?; 2) if so, how do these communities overlap
with one another? To answer these questions, we apply the popular OSLOM
(Order Statistics Local Optimisation Method) algorithm [
The remainder of this paper proceeds as follows. Section 2 describes related work on co-occurrence networks of the type used in this paper and community nding. Section 3 describes how the meetup network was constructed, how community nding techniques were applied to it, and how the resulting communities were labelled using textual metadata. Section 4 explores the network created and the communities found within it, focusing on a subset of largely technologyfocused communities. Finally, Section 5 concludes with suggested directions for future work in this area. 2 2.1
A co-occurrence network is a weighted network constructed such that each edge
indicates the frequency with which two items appear in the same context. In
some cases, this kind of network is formed by projecting an existing bipartite
network to a one-mode network by de ning the weights as the number of
cooccurrences. In other cases, such networks are created directly from raw count
data. For instance, in a physical co-location network, we might create an edge
between two individuals, where the edge weight indicates the number of times the
individuals were in the same place at the same time. Network analysis techniques
have been used to explore co-occurrence patterns in a range of domains. One
common application area has been in bibliometrics, where co-citation networks
have been analysed [
Interactive graph: https://draig.ucd.ie/MeetupNetDublinInteractive/
of co-occurrence networks include the analysis of word co-occurrence patterns
in natural language processing [
When analysing networks in many domains, we will often be interested in
performing community detection, where the goal is to identify the underlying group
structures in the data. Typically this is performed as an unsupervised task. A
substantial amount of work in this area has focused on the detection of disjoint
communities, where each node belongs to at most one community. Algorithms
in this context can be broadly grouped into three types:
(1) Hierarchical algorithms construct a tree of communities based on the network
topology. These can be one of two types: divisive algorithms [
In many real-world networks, we observe \pervasive overlap", such that
individuals frequently belong to many highly-overlapping communities [
This section describes how data was collected using the meetup.com API, how a meetup co-occurrence network was constructed, the application of community nding methods to this network, and the use of text analytics methods to label and explain communities found in the network. The Meetup.com website provides an open API2 that allows access to data from the meetup.com platform. Building a meetup co-occurrence network required data about meetup groups in Dublin and the users that are members of each group. The / nd/groups API call generates a list of all meetups in a speci ed country. Metadata providing the name, description, and host city of each meetup is also returned. Using this call we generated a list of all meetups in Ireland and then ltered this to exclude those not hosted in Dublin. Private meetups were also excluded. This left a ltered set of 1,482 Dublin-based public meetups. The /2/members API call generates a list of member IDs for all members of a speci c meetup group. We used this call to generate a list of all of the members of each Dublin-based public meetup group identi ed in the previous step. Motivated by the concept of co-citation networks in bibliometrics, we create an undirected weighted graph based on the member overlaps, or co-memberships, between pairs of meetups. Fig. 1 illustrates this approach. As Meetup 1 and Meetup 2 share common users, a link would exist between them in a co-occurrence graph, whereas neither of these would link to Meetup 3 as no common users exist.
In the proposed network, each unique meetup is represented as a node, and
a weighted edge between two nodes represents an association between the two
meetups represented by its endpoints. Every meetup has a set of registered
members, which allows us to measure the degree of overlap between the member sets
for pairs of meetups. Rather than looking at the raw number of common
members, we use a normalised value. Speci cally, the degree of association between
two meetups is calculated using the Jaccard index [
We apply community nding to the Dublin meetup network to organise it into a
smaller number of communities of related meetups that are easier to interpret, as
opposed to manually inspecting a large number of meetups individually. Given
that we might expect some users to be members of a diverse set of meetups, we
apply an overlapping community nding approach to the co-membership
network, which allows each meetup to potentially belong to multiple communities.
Speci cally, we apply the popular OSLOM algorithm [
OSLOM follows a greedy expansion strategy to detect communities by optimising a local tness function on seed nodes based on statistical measurement. This is done in three stages. In the rst stage, the algorithm searches for significant clusters by selecting a node at random as a seed then expanding it into a larger community. During this expansion, the algorithm tests the tness score of the neighbouring nodes of the seed node by using a statistical test that evaluates the signi cance of each neighbouring node to be added to the seed. In the second stage, for each community, the algorithm performs a process which involves removing or adding nodes to maximise its signi cance score. In the third stage, the hierarchical levels of the communities are detected. Due to the presence of a stochastic element during calculating a node's signi cant score the above stages are repeated several times and aggregated to create a stable set of communities.
To identify communities, we applied the undirected version of OSLOM to the meetup network using a range of values [0:01; 0:5] for the resolution parameter, which indirectly controls the size of the communities identi ed by the algorithm. After each run, we ltered communities containing < 5 nodes, on the basis that these do not represent signi cant groupings of meetups. Based on manual inspection, a value of 0:1 for the resolution parameter provided a good trade-o between ensuring that communities were coherent, while also ensuring that the number of duplicate communities (i.e. related to identical topics) was minimised. 3.4
To extract meaningful insights from the community nding results, and to examine the topical coherence of these communities, we produce human-interpretable labels for each community. This allows us to explain and understand the groups at a high level. Fortunately, the meetup.com API provides a rich set of metadata which can be used for this purpose. Speci cally, we propose a custom approach for labelling each community based on the short name eld and the longer description eld associated each meetup assigned to that community. Formally, we generate name labels for communities as follows: 1. For each meetup name eld, extract all alphanumeric terms and lter out common stopwords (e.g. \the", \meetup"', \group"). 2. Construct a meetup-term matrix A, such that each row corresponds to a meetup, each column corresponds to a term, and each entry indicates the number of times a term appears in a meetup name. 3. Apply standard log-based TF-IDF weighting to the matrix, and normalise the rows to unit length to account for di erent name lengths. 4. For each community C: (a) From A, compute the mean vector of the rows corresponding to the meetups which have been assigned to C. (b) Rank the values in the mean vector in descending order, and select the top t terms to create a name label.
We apply an analogous procedure to generate description labels for communities, based on the longer meetup description text. 4
In the original co-membership graph, each meetup group is represented by a node and the edges indicates the weight of the connection between pairs of meetups. The network consists of 1,482 nodes connected by 1,416,326 weighted edges, which represents a single connected component. 4.1
As is typical of many co-occurrence networks, the Dublin meetup network is highly dense, with edges present between 64.5% of all possible pairs of nodes. However, there is some variation in terms of the weights on these edges. We observe that 86.63% of the weighted edges have values 0:01. This is indicative of a relatively small intersection between the memberships of many meetups.
When ranked by number of members, unsurprisingly, the three largest
meetups are from the social sphere: New and Not So New In Dublin (21,149
members), Events, Drinks and Dancing in Dublin (16,582 members) and Dublin
International (14,812 members). It is more informative, however, to measure the
importance of meetup nodes in the overall meetup network and we do this using
centrality analysis. A range of measures have previously been proposed for this
task. We focus on two popular measures of centrality which are designed for use
on weighted networks:
1. In weighted eigenvector centrality, a node is deemed more important if it is
connected to other important nodes. In the weighted variant of this measure,
centrality scores are calculated based on the rst left eigenvector of the
weighted graph adjacency matrix [
The top 10 meetups as ranked by each measure are listed in Table 1. It is interesting to note the signi cant di erence between the lists generated by these two approaches|in fact there are no meetups that occur in both lists. The dominance of technology related meetups in Table 1a re ects the vibrance of the tech community in Ireland and indicates a slight bias on the meetup.com platform towards technology savvy users. It also suggests the existence of a cluster of similar meetups attended by a core group of overlapping members. The meetups in Table 1b re ect the fact that betweenness centrality measures the ability of nodes in a network to connect disparate parts of that network. Although there are some technology related meetups here, most focus on topics of broad appeal (e.g. Speak English Dublin) that are likely to attract members with disparate other interests. 4.2
The application of the OSLOM community nding algorithm yielded a model with 26 communities, ranging in size from 17 to 216 meetups. The mean community size was 65. Table 2 summarises details for the 15 largest communities identi ed in the normalised meetup co-membership network (full list is available in 1). For each community, we report its size (i.e. number of meetups) and labels generated based on meetup name and description metadata, using the Rank Meetup Name 1 Dublin Arti cial Intelligence & Deep Learning 2 Big Data Developers in Dublin 3 Data Scientists Ireland 4 Zalando Tech Events Dublin 5 Machine Learning Dublin 6 Data Science and Engineering Club 7 Hackers and Founders Dublin 8 GDG Dublin 9 Dublin Startup Founder 101 10 Dublin - Coder Forge
(a) Weighted eigenvector centrality Fig. 2: A visualisation of the meetup network for Dublin, where only intracommunity edges are shown. Each node represents a meetup, where its size is proportional to the number of members in the meetup. Darker colour indicate a higher weighted degree of the nodes, numbers indicate community Ids. approach described in Section 3.3. These labels indicate a diverse range of communities, covering topical areas such as technology, travel, self-help, music, and entrepreneurship. Despite the fact that the name and description elds typically di er considerably in length, it is interesting to note the relatively high level of overlap for the terms appearing on both lists for the same community.
To further explore the results produced by OSLOM, for each community we constructed a subgraph of the original network, and then ranked the nodes assigned to their community based on their centrality within that induced subgraph, as calculated by weighted degree centrality. The score for a node i is de ned as the sum of the weights of the edges connecting i and its neighbours. Table 3 lists the top 3 most central meetups in each of the 15 largest communities.
These 26 subgraphs are shown in Fig. 2 generated using the MultiGravity
Force Atlas 2 graph layout algorithm provided by the Gephi tool [
The level of overlap between communities was not as high as might be expected. In total, 197 of the 1,482 meetups were assigned to more than one comDescription Label fun, members, friends, time, hikes, free, social, friendly, looking, food healing, life, meditation, experience, self, energy, practice, spiritual, mind, mindfulness data, programming, developers, community, code, science, software, technology, technologies, learn data, learn, share, learning, developers, cloud, community, security, technology, software business, marketing, digital, entrepreneurs, startup, market, network, owners, sales, job yoga, life, body, meditation, class, health, classes, practice, energy, mind Name Label hiking, international, wicklow, friends, yoga, book, culture, adventure, language, travel meditation, yoga, healing, spiritual, heart, sound, empowerment, soul, life, positive data, user, science, tech, engineering, big, cloud, users, things, learning user, tech, security, cloud, sharepoint, technology, game, software, data, crypto business, digital, marketing, startup, entrepreneurs, network, job, professionals, innovation, market yoga, meditation, workshop, stress, dun, laoghaire, camino, running, dance, therapy startup, entrepreneurs, digital, lean, business, entrepreneurs, marketing, business, marketing, agile, growth, prod- startup, networking, digital, lean, uct, innovation product, community, innovation yoga, health, happiness, meditation, ve- yoga, life, meditation, help, support, gan, prayer, empowerment, circle, cen- healing, learn, world, health, work tre, self user, mysql, traders, developers, tech, js, learn, product, developers, mysql, product, data, sprint, net share, community, meetups, professionals, technologies, engineers music, night, friends, fun, singles, rock, singing, love, members, sing yoga, meditation, body, classes, life, mind, healing, health, practice, nature music, singles, rock, social, travel, south, international, fans, electronic, 30s yoga, meditation, health, healing, classes, relaxation, self, body, light, sound empowerment, self, book, support, health, workshop, eating, therapy, life, development circle, things, drinks, city, fun, hike, ladies, social, friends, book dance, dancing, yoga, classes, movement, salsa, tness, class, set, handstand soul, prayer, network, life, healing, workshop, empowerment, biodanza, centre, body life, world, diet, work, feel, learn, share, spiritual, ideas, nd drinks, friends, women, fun, book, food, wants, single, cinema, dinner dance, classes, dancing, fun, tness, workout, 8pm, levels, class, movement life, god, healing, faith, spiritual, love, evening, work, reiki, chat 4 7 17 14 22 3 25 10 18 8 21 15 16 26
Size 216 148 137 118 84 80 78 77 71 63 61 54 53 53 52 munity. Of these, 176 appear in two communities, 20 in three communities, and a single meetup appears in four communities (Headless Awareness Dublin ). Table 3 also reports the percentage of overlapping nodes in each of the 15 largest communities|i.e. how many nodes assigned to the community were also assigned to at least one other community. For some communities, a reasonably high proportion of nodes are overlapping (e.g. communities 21, 22, 25 in Table 3), while in other cases the vast majority of nodes belong exclusively to that community (e.g. communities 7, 10, 15 in Table 3).
To further illustrate the value of community nding, we present a more detailed analysis of a small subset of seven of the 26 communities found that have been manually identi ed as relating to technology. This subset is highlighted in 4 7 17 14 related to sports. The overlaps between these communities are also interesting. For example, community 12 has a relatively large overlap with community 1, and community 9 has a relatively large overlap with community 6. 5
This paper demonstrated the use of network analysis techniques to explore public data collected from the meetup.com platform, to characterise the fabric of a city. A co-membership network of meetups from Dublin, Ireland was constructed. This network allowed us to reveal the most important meetups in Dublin via measures of node centrality. To uncover the structure of the network at a higher level, community nding techniques were applied. By subsequently applying text analysis procedures to the aggregated metadata associated with each community, we have shown that thematically-coherent communities exist within Dublin's Meetup ecosphere. We also observed a limited degree of overlap between certain communities, where users might naturally share common interests.
Although the present study speci cally focuses on Dublin's meetup network, using the same framework, the underlying communities of other cities could also be explored, and future work will develop a tool to support this. It would also be interesting to incorporate additional layers of meetup metadata into the network construction process. For example, some meetup groups are much more active than others, as some users are much more active than others, and this could be used to lter nodes in the network, and to in uence the weight of edges.