1.1. Background
We explore the application of social network analysis and computational content analysis towards studying the motivations and behaviours of members of digital sharing communities located in three towns in Singapore. We present the process of data collection, as well as social network analysis to extract community structures and member relationships, and computational content analysis to discover dominant topics of conversation as well as topics that lead to offline neighbourhood engagement.
1.1.1. Sharing Neighbourhoods
With more than 1 million flats that are under HDBās management, HDB has committed to fostering social cohesion and social vibrancy within neighbourhoods. To do so, HDB had previously encouraged the development of sharing communities through initiatives such as Hello Neighbour! @ Tampines Central and the Good Neighbours Movement.
1.1.2. The Rise of Online interactions
The rise in digitisation and online platforms of communication has led HDB to observe digital space as new grounds for the development of these sharing communities. Sharing communities have been observed to move some of their activities from physical space to digital space, while those in digital space may have effects on physical engagement.
1.2. Digital Sharing Communities
Thus, we define Digital Sharing Communities to refer to communities that exist on online social media platforms, but also associated with a place (region, town, block), that not only share physical resources but also include knowledge sharing and social sharing.
1.3. Research Questions
How can we define characteristics to describe the various digital sharing communities?
Do digital sharing communities have potential to generate neighbourhood engagement?
Authorship
This study was conducted by Ivan Chuang, Asha Suresh, Gong Hailun & Huang Yimin, as a Research Studio for the Master's in Urban Science, Planning and Policy programme (MUSPP) @ Singapore University of Technology and Design (SUTD). 2021.
2.1. Rationale
Our rationale for this data collection was to sample examples of digital communities to construct a typology of digital communities, by uncovering the common and distinguishing characteristics between the variety of digital communities in the wild. Through this typology, we can then discover the factors that contribute to a vibrant and active digital community.
2.2. Method
Our data collection involved the collection and curation of a list of sharing communities within neighbourhoods that we observed on digital platforms. After this initial data collection, we categorized and analyzed the groups.
We conducted our data collection on the digital sharing communities centred within three towns in Singapore: Punggol, Toa Payoh, and Jurong. The reason we chose these towns was due to their estate maturityāPunggol is considered a young town, Jurong a middle-aged town, while Toa Payoh is often considered an older town. We compare between these three towns to discover if the maturity of an estate has an effect on digital communities that belong to these estates.
Our selection criteria for a digital sharing community to belong to a town informed our search process to discovering these communities online. We consider a digital sharing community to belong to a town if it has one or more of these characteristics:
To discover these digital communities, we used a combination of keyword search and snowball sampling to search for these communities online. The main digital platforms search bars that we used were Facebook, Instagram and Google.
2.2.1. Keyword Search
Keyword searches are defined as a method of searching for information which possess keywords specified by a user, whereby a search term list is applied to a full text index to find documents containing one or more words that are specified by a user (Chen et al, 2020). For our keyword search, we used location names at three levels of place scales we had specified, as well as their combined names:
During actual application, additional keywords were added to allow for better searches. These additional keywords aligned with the group purpose, such as āgroup buyā, āinterest groupā or āBTOā. We felt that applying a combination of group place scale coupled with potential group purpose gave better search results in the form of sharing communities that fit our criteria.
2.2.2. Snowball Sampling
In traditional snowball sampling, researchers ask informants to recommend additional contacts who might be able to contribute to a study, thus partially reflecting the organic structure and development of social networks (Baltar, 2012).
In our study, we used a modified form of snowball sampling in the digital space to reflect the complexities of digital networks recommendations and linked web pages. This took into account sharing communities that were not as discoverable on digital platforms, such as communities with unconventional group names that did not align with our keyword searches.
For example, if we entered a group on Telegram and noticed that a user had shared a link to another sharing community in the group, we would click on sent link to see if the suggested group met our criteria (namely a neighbourhood sharing community from either Punggol, Toa Payoh or Jurong that was on a digital platform). If the suggested group met our criteria, we would include it in our data collection.
Another instance where snowball sampling came in useful would be a trend that we observed amongst some sharing communities, where the digital presence of these groups straddled multiple platforms. An example was sharing communities based on instagram that included a link in their group description to the shared Whatsapp group where the actual transactions occurred. Namely, Whatsapp was the primary digital basis of interaction, while Instagram was a secondary medium. In these instances, snowball sampling proved useful in enabling us to identify sharing communities that we may not have found solely via keyword searches.
2.3. Results
2.3.1. Distribution Across Towns
In total, we found 81 instances of digital sharing communities between the three towns.
We observed a positive correlation between the number of digital sharing communities and the maturity of the town, as seen from Figure 1. Namely we observed that the youngest town, Punggol, had the highest number of digital sharing communities (n=37), as compared to middle aged town Jurong (n=26), and old town Toa Payoh (n=18).
Given our method of searching, we cannot confirm that these are the exact proportions of sharing communities within each neighbourhood. Rather, we can claim that the sharing communities in younger towns are most easily searchable and found on digital platforms.
This could be attributed to the higher digital literacy of Punggol residents, given their younger demographic makeup, with 22% of residents aged 30-39 years old, comparing to 13% in Toa Payoh and 14% in Jurong. (Singapore Department of Statistics, 2020) This could also be due to the fact that there is higher focus on digital within newer towns, seen by the focus on developing Punggol as Singaporeās first digital district (URA, 2020). This aligns with broader goals of Singaporeās goal to be a globally recognized leading smart city with a strong focus on digital innovation and technology (Yee, 2019). Towns that were more recently redeveloped thus are more intertwined with technology perhaps as compared to older towns, meaning that residents interact with and use technology more in their day to day lives such as the method in which they reach out to like-minded individuals for community creation.
2.3.2. Distribution Across Platforms
Figure 2 breaks down the sharing communities found in each of the different towns into which platforms the groups were discovered on. The predominant platforms include Whatsapp, Telegram, Instagram and Facebook.
Among the social media platforms, Facebook and Telegram emerge as platforms that best facilitate the discovery of digital sharing communities. A significant proportion of communities were discovered on Facebook amongst the three towns: Punggol (n=17), Toa Payoh (n=7), and Jurong (n=7).
We noted that WhatsApp communities were the least discoverable as the platform does not offer a search function, and entry into the community requires invite links. Only six communities were found from Punggol (n=4), Toa Payoh (n=1), and Jurong (n=1). The communities we did find were found via snowball sampling, as they had promoted their invite links on the other social platforms. We observed that instagram channels were popularly used as platforms for the promotion of invite links to other platforms. This is due to our observation that community accounts on Instagram rarely used the platform for community engagement, and instead used it as an advertising medium to direct people towards their actual community hosted on another social platform.
By observing the age of group creation, we can observe a trend in platform preferences. This is seen in Figure 3, where based on the year of creation, Facebook was the popular platform for many initial digital sharing communities from 2008 to 2014. Afterwhich other platforms such as Instagram and Whatsapp came into the picture and more community groups were formed there. We observed that Telegram has been gaining popularity in recent years, as seen from 2017 onwards where the number of communities formed on Telegram rose more as compared to other platforms.
2.3.3. Purposes of Digital Sharing Communities
The distribution of these different communities across the three towns can be seen in Figure 4. We discovered 5 classifications of purposes of digital sharing communities, which we have classified as: Residential, BTO, Interest Group, Groupbuys and Community Markets.
The distribution of these different communities across the three towns can be seen in Figure 5.
Residential (Punggol, n=14 ; Jurong, n=11 ; Toa Payoh, n=11) and Groupbuys (Punggol, n=11 ; Jurong, n=7 ; Toa Payoh, n=4) were the most frequent communities observed, a characteristic observed in all three towns. This makes sense as these communities are relevant to the largest proportion of residents, while BTO (Punggol, n=6 ; Jurong, n=2 ; Toa Payoh, n=2) and Interest Groups (Punggol, n=5 ; Jurong, n=3 ; Toa Payoh, n=4) communities target a smaller demographic based on the specific common topic of the group.
By looking at when these communities were created along the years, we can also observe that while Residential communities have been around for a while, Groupbuys are a more recent phenomenon, experiencing sharp growth in 2020.
2.3.4. Association between Place Specificity and Community Size
Place specificity defines the physical boundary of the location associated with the community. We discovered that the communities found could grouped into three levels of place specificity:
When we break down the communities by their place specificity, i.e. the physical bounded location that is associated with the community, we see that Punggol has the highest number of estate-level communitis, while Toa Payoh and Jurong have more town-level communities.
One possible reason for this is that the younger demographic of Punggol residents might prefer smaller communities. Small communities are easier to manage and more capable of fostering stronger neighbourly ties.
However, when we look at the mean size of these communities, it becomes apparent that generally, the size of the community scales as it is associated to a broader location. This makes sense as more residents would identify with the location associated with the community. Hence, region-level communities are able to reach a wider group of people, while estate-level communities are smaller and more diverse.
2.3.5. Association between Accessibility and Place Specificity
We were also able to categorise our communities by accessibility using 4 categories:As seen in Figure 8, we find that there are most open communities on town level, though estate has many open communities, it may be thanks to its big cardinality. In addition, Estate level has most communities which are both closed with permission and verification. According to this finding, we can assume that smaller scales are more likely to have closed communities because their social circle may be smaller and more closed.
The mean size of the communities justifies our previous assumption. We observe that the mean size of open communities on estate level is about half of the closed with verification communities, though there are more open ones. And on town and region levels, we can see a trend that communities in large scale are more likely to open because their target users are in a bigger range and permission or verification will be the restriction to let people join in. In conclusion, the differences in accessibility are mainly related to the groupsā aim and users.
2.3.6. Association between Purpose of Community and Organisation Structure
The types of organization also implicates the characteristics of a particular community. We categorised our communities into 3 categories by types of organization:
We can easily see the effect of type of organization on the choice on platform and the purpose of the communities. For different purposes they have different preferences of interactions, BTO and Interest Group Communities have more open discussions, while Groupbuys and Resident Communities are more like announcement channels. Additionally, most of the communities have flat hierarchy organization possibly because they want to encourage interactions among the members, which can have an effect on enhancing the social cohesion.
References
[1] Chen, Lisi, Shang, Shuo, Yang, Chengcheng, & Li, Jing. (2020). Spatial keyword search: A survey. GeoInformatica, 24(1), 85-106.
[2] Baltar, Fabiola, & Brunet, Ignasi. (2012). Social research 2.0: Virtual snowball sampling method using Facebook. Internet Research, 22(1), 57-74.
[3] Singapore Residents by Planning AreaSubzone, Age Group, Sex and Type of Dwelling, June 2000-2020. (2020). Singapore Department of Statistics. (Source)
[4] Urban Redevelopment Authority (URA). (2020d). Punggol Digital District: The Next Generation Smart & Integrated District. (Source)
[5] Yee, Y. W. (2019). Singapore is worldās smartest city: IMD Smart City Index. Straits Times. (Source)
3.1. Rationale
Our rationale behind conducting a social network analysis is to extract structures of digital communities and characterise them quantitatively, as well as characterise users within the community based on their interactions.
3.2. Method
3.2.1. Selection Criteria
Data availability is a necessity for social network analysis. The only platform that satisfied this necessity was Telegram. On the Telegram platform, BTOs and Residential digital communities are the most prevalent type of digital community. Hence, we selected a BTO and Residential community for each of the three towns, providing us six communities for study through social network analysis.
The criteria for the 6 selected groups is as following:
The selected groups can be seen in Figure 11.
# | Town | Purpose | No. of users | Year Created |
---|---|---|---|---|
1 | Punggol | Residential | 1217 (Large) | 2019 |
2 | Punggol | BTO | 561 (Medium) | 2019 |
3 | Jurong | Residential | 489 (Medium) | 2017 |
4 | Jurong | BTO | 265 (Medium) | 2019 |
5 | Toa Payoh | Residential | 220 (Medium) | 2020 |
6 | Toa Payoh | BTO | 73 (Small) | 2020 |
3.2.2. Network Analysis
We used the open-source GEPHI network visualisation software to analyse our social networks. To discover community structures within our social networks, we partitioned the users using modularity classes. Modularity produces clusters by looking for nodes that are more connected together than the rest of the network. We also measured the out-degree of each node to identify active users who engage with a wider proportion of their community.3.3. Network Visualisation
Legend
community member
color: modularity cluster
text message
add member
remove member
thickness: no. of interactions
3.4. Findings
A network analysis of a social community allows us to define quantitative metrics that measure the interactivity of each member within the network. We calculate these three metrics for each member in the network:
With this, we characterise the members within the six selected digital sharing communities:
3.4.1. Structural Forms of Communities
We found that communities generally take three forms.
3.4.2. A typology of users
3.4.2.1. Super Members
We identified a type of member called a super member which is characterised by a high degree relative to the other members in the community. Super members were found within 3 of the digital sharing communities: Punggol BTO, Punggol Residential and Jurong BTO, as seen in Figure 13.
3.4.2.2. Lurkers
We observed the presence of many members within that do not interact with other members (degree=0), which we have labelled as lurkers. The high proportion of lurkers is characteristic of open groups, which do not exclude members based on their activeness. However, because we use the explicit mention of another user to measure interaction (such as tagging a user i.e. '@user123'), our method is limited in that it cannot detect interactions that happen without explicit mentions.
3.4.2.3. Bridging Members
Within the chat groups in Jurong, we identified members that have high betweeness centrality while having low degree. These are members of the community that while do not interact as often as super members, are integral in connecting sub-communities found within the larger community. These sub-communities themselves can be found through modularity. (Modularity clustering is visualised by the node color in Figure 12.)
The presence of bridging members is also a measure of how cohesive a community is. In the Punggol communities, we observe both the lack of bridging users, as well as the identification of one majority cluster by modularity.
4.1. Rationale
Our rationale behind performing a content analysis on the text messages within these communities is to discover the topics that 1) anchor members to the community and 2) lead to offline neighbourhood engagement.
4.2. Methodology
4.2.1. Pre-processing
To perform a computational content analysis, the text messages were first tokenised and lemmatised using the Spacy NLP library. Stopwords were also removed. This reduced the complexity of our unit of analysis.
4.2.2. Topic Modelling via Entity Recognition
To identify the most common topics in each community group, we assume a topic to be composed of multiple Named Entities: real-world object, such as persons, locations, organizations, products. Hence, we apply a Named Entity Recognition (NER) model using the NLP Spacy library to extract entities from the text messages for each community.
We then count the frequency of occurrence of each entity, and identified 5 common topics across all six communities, which we code to the entities.
4.2.3. Keyword-in-Context
To measure the instances of messages that relate to offline neighbourhood engagement, we filtered for text messages after lemmatisation that fulfilled the following query:
contains(āmeetā OR āgoā OR ācomeā) AND contains(āweā OR āusā OR āyouā)
The filtered messages were then manually verified for content.
4.3. Results
4.3.1. Community Topics
5 codes emerged from the analysis of named entities, which we interpret as topics of which these entities belong to:
Here we visualise the topic distribution across the six groups.
Entities | Counts | Category |
---|---|---|
(HDB, ORG) | 544 | Housing |
(WWSR1, LOC) | 73 | Location |
(656a, LOC) | 68 | Location |
(HLE, ORG) | 58 | Housing |
(BTO, ORG) | 58 | Housing |
(Jackson, PERSON) | 54 | Housing |
(Punggol, LOC) | 49 | Location |
(WWSR2, LOC) | 46 | Location |
(Singapore, GPE) | 40 | Location |
(CPF, ORG) | 37 | Housing |
Entities | Counts | Category |
---|---|---|
(HDB, ORG) | 169 | Housing |
(Punggol, LOC) | 58 | Location |
(BTO, ORG) | 38 | Housing |
(NSR, LOC) | 28 | Location |
(Singapore, GPE) | 28 | Politics & Current Affairs |
(HLE, ORG) | 27 | Housing |
(OA, OBJ) | 60 | Housing |
(Daikin, ORG) | 23 | Amenities & Services |
(WhatsApp, ORG) | 46 | Community |
(Aircon, OBJ) | 14 | Amenities & Services |
Entities | Counts | Category |
---|---|---|
(Singapore, GPE) | 159 | Politics & Current Affairs |
(Hong Kong, GPE) | 142 | Politics & Current Affairs |
(China, GPE) | 105 | Politics & Current Affairs |
(SDP, NORP) | 69 | Politics & Current Affairs |
(Parliament, ORG) | 66 | Politics & Current Affairs |
(Chinese, NORP) | 62 | Community |
(BB, LOC) | 60 | Location |
(Singaporeans, NORP) | 59 | Community |
(GE, EVENT) | 46 | Politics & Current Affairs |
(PAP, NORP) | 39 | Politics & Current Affairs |
Entities | Counts | Category |
---|---|---|
(Chinese, NORP) | 34 | Community |
(CJ, PERSON) | 28 | Community |
(Wolflet, PERSON) | 28 | Community |
(Singapore, GPE) | 26 | Politics & Current Affairs |
(10, MONEY) | 21 | Amenities & Services |
(Clementi, LOC) | 19 | Community |
(MRY, PERSON) | 15 | Community |
(Jurong, LOC) | 15 | Location |
(Joyce, PERSON) | 15 | Community |
(Jim, PERSON) | 15 | Community |
Entities | Counts | Category |
---|---|---|
(Bartley, LOC) | 5 | Location |
(HLE, OBJ) | 3 | Housing |
(Bartley GreenRise, LOC) | 2 | Location |
(Toa Payoh, LOC) | 5 | Location |
(Optional Component Scheme, OBJ) | 2 | Housing |
(900k, MONEY) | 2 | Housing |
(Kim Keat, LOC) | 2 | Location |
(Bidari, LOC) | 2 | Location |
(Blk 650, LOC) | 1 | Location |
(Biddari, LOC) | 1 | Location |
Entities | Counts | Category |
---|---|---|
(50, MONEY) | 27 | Amenities & Services |
(60, MONEY) | 26 | Amenities & Services |
(5, MONEY) | 25 | Amenities & Services |
(ChewApp, ORG) | 23 | Amenities & Services |
(30, MONEY) | 23 | Amenities & Services |
(7, MONEY) | 23 | Amenities & Services |
(Singapore, GPE) | 20 | Location |
(WhatsApp, ORG) | 19 | Community |
(Poseidon, ORG) | 18 | Amenities & Services |
(Grounded Pleasures, ORG) | 18 | Amenities & Services |
We discovered that each community had a dominant topic whose entities make up more than 50% of identified entities for each community, with some communities having one or two secondary topic. Each community had at least three topics.
Location | Type | Amenities & Services | Community | Housing | Location | Politics & Current Affairs |
---|---|---|---|---|---|---|
Punggol | BTO | 2.72% | 7.35% | 56.73% | 29.10% | 2.10% |
Punggol | Residential | 11.59% | 2.99% | 51.96% | 28.22% | 5.23% |
Jurong | BTO | - | 16.55% | 2.64% | 8.36% | 72.45% |
Jurong | Residential | 21.26% | 54.19% | - | 16.77% | 7.78% |
Toa Payoh | BTO | - | 1.96% | 62.75% | 35.29% | - |
Toa Payoh | Residential | 90.13% | 3.81% | - | - | 5.06% |
A community's name does not necessarily reveal the dominating topic within a community. In the Jurong BTO community, politics and current affairs was the dominant topic, with only 2.6% of identified entities related to housing. In the Toa Payoh residential community, amenities and services was the dominant topic (90.3%) due to the prevalance of messages related to groupbuys.
4.3.2. Potential for Neighbourhood Engagement
We found that messages in BTO and Residential groups rarely lead to offline neighborhood engagement, contributing to a tiny portion of the total messages.
Location | Type | Total Messages | Count | False Positives | True Positives | % of True Positives over Total Messages |
---|---|---|---|---|---|---|
Punggol | BTO | 18107 | 150 | 137 | 13 | 0.1% |
Punggol | Residential | 9373 | 69 | 55 | 14 | 0.2% |
Jurong | BTO | 3358 | 41 | 22 | 19 | 0.6% |
Jurong | Residential | 10558 | 57 | 29 | 28 | 0.3% |
Toa Payoh | BTO | 558 | 2 | 2 | 0 | 0% |
Toa Payoh | Residential | 377 | 10 | 9 | 1 | 0.3% |
Our attempt at an automatic filtering also revealed the limitations of such an approach, and that there is still a need for manual verification.
We also traced these messages that lead to offline neighbourhood engagement back to their users and discovered that while super-members were not the majority, in many instances, super-members were the initiators of the message chain related to neighborhood engagement.
Looking at contents of the messages, we discovered that while motivations behind meeting up varied across the communities, socialising was a common factor.
Location | Type | No. of Messages | No. of Unique Actors | No. of Super-members | Topic |
---|---|---|---|---|---|
Punggol | BTO | 13 | 9 | 2 | Socialising |
Punggol | Residential | 14 | 12 | 2 | BTO-related |
Jurong | BTO | 19 | 8 | 0 | Organised Family Activities |
Jurong | Residential | 28 | 8 | 2 | Socialising / Pokemon Go |
Toa Payoh | BTO | 0 | 0 | 0 | - |
Toa Payoh | Residential | 1 | 1 | 0 | - |
5.1. Summary of Findings
5.2. Conclusion
5.2.1. Methods for Studying Digital Sharing Communities
It is important that we study Digital Sharing Communities as they are new generators of neighborhood engagement. Studying the existing communities can inform us on how we can promote more of these communities, or even envision completely new forms.
However, Digital Sharing Communities vary in form and function and each demands different methods for studying them. We have shown how several of both quantitative and qualitative methods can be used to gain insight on these communities. The primary topics and activities of a group cannot be based on their named intentions alone but requires a deeper analysis of the interactions exchanged in the group itself.
5.2.2. Potential for Neighbourhood Engagement
Interest groups and group buys show more potential for neighbourhood engagement. This is because neighborhood engagement is centred on recurring activities that people can commit to and set aside time for.
However, this is not the case for interest groups and group buys. Instead, most neighborhood engagement took the form of spontaneous, impromptu, and informal meetups. There is no one āin-chargeā of fostering neighborhood engagement.
In all healthy communities observed, there is at least one person that anchors the rest of the other members in the group. For residential and BTO groups, these are the super-members, not necessarily the admins. For Group Buys and Interest Groups, these are the admins themselves. In both communities, admins and super-members are the primary initiators of digital interaction that leads to neighbourhood engagement.