Usecases/Clustering

Clustering Usecases

In social networks, clustering is used to find groups of people with similar likes and dislikes in an effort to predict future behavior. Here are some uses cases:

Predicting Telecom Churn

In a society, an individual's behavior can be heavily influenced by that individual's close friends. Examples of such behaviors include but are not limited to alcohol consumption, health related habits, like/dislike products, and adopting/rejecting services. Specifically, in the telecommunication domain, there is a lot of churn and service providers pay a lot of attention to spotting potential churners. To service providers, keeping existing customers is as important as gaining new customers. One way to identify potential churners is to 1) model the underlying social network and call network as a graph with attributes describing users and their interactions, 2) run graph analytics against the graph to identify communities, and 3) rank existing users based on the attrition rate of their close friends in a community, and possibly other signals.

Features: graph clustering, vertex ranking

References:

[1] The Influence of Parents and Friends on Adolescent Substance Use: A Multidimensional Approach http://www.ncbi.nlm.nih.gov/pmc/articles/PMC3132133/pdf/nihms294477.pdf

[2] Collective Churn Prediction in Social Network http://www.mysmu.edu/faculty/fdzhu/paper/ASONAM%2712.pdf

[3] Local Graph Sparsification for Scalable Clustering http://www.cse.ohio-state.edu/%7Esatuluri/satuluri_sigmod11.pd

[4] Statistical Properties of Community Structure in Large Social and Information Networks http://cs.stanford.edu/people/jure/pubs/ncp-www08.pdf

[5] Detecting community structure in networks http://www-personal.umich.edu/~mejn/papers/epjb.pdf

Identify influencers/experts/expertise

Finding the right people is oftentimes a key to an efficient resolution of real-world problems, technical and business alike. It is very useful to be able to identify accurately subject matter experts in a social media or internal site (e.g. blogs, discussions, forums, etc). One approach is to rely on end users' feedback and ranking. This is effective but obviously a manual process. An automatic approach is to run graph analytics against the underlying social graph formed by the users and their interactions, and identify influencers by calculating metrics like page ranking, centrality, etc.

See also presentation from Boeing at the W3C Social Business Workshop that discusses this and related usecases.