Recommending Users: Whom to Follow on Federated Social Networks
| Key | Value |
|---|---|
| Author | Trienes et. al |
| Year | 2018 |
| Link |
- Problem: Recommendation algorithms important for community building but hard in DOSNs
- Solution: Proposes two user recommendation algorithms and test it on Mastodon data
- Interesting takes on data collection, important for own collection: robots.txt
# Extracted notes
social networks employ recommendation algorithms that filter large amounts of contents and provide a user with personalized views of the network [@WZI6JC4Z#Trienes_Cano_Hiemstra_2018, p. 1]
recommendations by listing profiles a user may be interested to connect with [@WZI6JC4Z#Trienes_Cano_Hiemstra_2018, p. 1]
large-scale user-surveillance and the miss-use of user data to manipulate elections [@WZI6JC4Z#Trienes_Cano_Hiemstra_2018, p. 1]
decentralizing authority and promoting privacy [@WZI6JC4Z#Trienes_Cano_Hiemstra_2018, p. 1]
networks more attractive and promote community building, we investigate how recommendation algorithms can be applied to decentralized social networks [@WZI6JC4Z#Trienes_Cano_Hiemstra_2018, p. 1]
collaborative filtering recommender based on BM25 and a topology-based recommender using personalized PageRank [@WZI6JC4Z#Trienes_Cano_Hiemstra_2018, p. 1]
aim to promote user control by decentralizing authority and relying on open-source software and open standards [@WZI6JC4Z#Trienes_Cano_Hiemstra_2018, p. 1]
traditional social media, one key success factor of such a network is an active and engaged community. [@WZI6JC4Z#Trienes_Cano_Hiemstra_2018, p. 1]
As a community grows, overwhelming amounts of content make it increasingly difficult for a user to find interesting topics and other users to interact with. For that reason, popular platforms such as Twitter, LinkedIn and Facebook introduce recommender [@WZI6JC4Z#Trienes_Cano_Hiemstra_2018, p. 1]
systems that set out to solve a particular recommendation task. [@WZI6JC4Z#Trienes_Cano_Hiemstra_2018, p. 1]
MetropolisHastings Random Walk (MHRW) adapted for directed graphs [@WZI6JC4Z#Trienes_Cano_Hiemstra_2018, p. 1]
unbiased sample of the Mastodon user graph. [@WZI6JC4Z#Trienes_Cano_Hiemstra_2018, p. 1]
Acquiring the complete graph of a social network is always infeasible due to API limits and time constraints [13]. An additional concern arises in a distributed social network. As data is not stored at a central authority, there is no single API that provides access to all parts of the network. Instead, data is scattered around different sub-networks [@WZI6JC4Z#Trienes_Cano_Hiemstra_2018, p. 2]
To overcome the time constraint, we apply the Metropolis-Hastings Random Walk (MHRW) to acquire an unbiased sample that is still representative of the complete graph. [@WZI6JC4Z#Trienes_Cano_Hiemstra_2018, p. 2]
Due to the fact that a distributed social network has no central API, one has to query the API of each individual sub-network referred to as instance. [@WZI6JC4Z#Trienes_Cano_Hiemstra_2018, p. 2]
During the data collection, we apply fair crawling policies. Only instances that allow crawling as defined by the robots.txt are considered. [@WZI6JC4Z#Trienes_Cano_Hiemstra_2018, p. 2]
when parts of the network are considered to be out of date (i.e., when the cache expires). Second, and more importantly, such an approach seems to be in conflict with the intentions behind decentralization. By constructing a database that aims to capture the entire network graph, one starts to centralize the data of a federated social network. [@WZI6JC4Z#Trienes_Cano_Hiemstra_2018, p. 4]
to reduce the overhead associated with crawling in an online setting, one might attempt to gradually construct a cached representation of the entire network graph. [@WZI6JC4Z#Trienes_Cano_Hiemstra_2018, p. 4]
Acquiring a sufficiently large snapshot of the network topology for offline recommendation proofed to be difficult and costly. Keeping the snapshot up-to-date needs constant re-sampling. Online recommendation was done by sampling the graph neighborhood for the current user. [@WZI6JC4Z#Trienes_Cano_Hiemstra_2018, p. 4]
First, studying the extent to which incomplete data impacts the recommender performance may derive methods that are tailored towards federated social networks which operate with limited amounts of data. Second, user recommendation algorithms in popular social media increasingly utilize user context information such as location data and interests. It remains unclear how such data can be effectively acquired and utilized in federated social networks while preserving privacy. Third, BM25 might not be the best ranking function for the presented recommender approach, and it should be compared to functions that also use popularity-based scoring. Finally, one may investigate how decentralized communication protocols such as ActivityPub can be extended to support community building algorithms while maintaining the notion of decentralized network data. [@WZI6JC4Z#Trienes_Cano_Hiemstra_2018, p. 4]