By Shafi Rahman
The idea of social influencers is quite enticing for marketers. Mass marketing is often an expensive and blunt tool, and its impact is usually very difficult to assess. Peer-to-peer, word-of-mouth publicity is sometimes considered an effective alternative for targeting and generating desired outcomes. These marketing campaigns can be made more effective if they are seeded with individuals with high social influence, making a larger impact with lower costs.
Social network platforms have opened up new opportunities for identifying social influencers. There are algorithms such as graph theory to measure node centrality and node relevance in a network. They can be applied on social networks with context specific influence metrics to identify the nodes (individuals) that are most central or relevant. The challenge is to identify whether these individuals are truly influencers, or the folks that they are supposedly “influencing” are merely similar to them. Thus the problem is how to identify the confounding factors and isolating treatment effects of true influencers.
For a data scientist a randomized test is the touchstone for addressing confounding factors and measuring treatment effects. Social media platforms provide fertile and effective ground for setting up such tests. In a recent article published in Harvard Business Review, the authors set up randomized tests on one such social media platform. A hold out group was created and then multiple test cells were set up with different treatments to test the extent of impact of various treatments.
Such similar tests could be repeatedly administered on various small test cells (micro-cells) to identify true social influencers. Unfortunately, these tests are usually not easily scalable due to the need for expert intervention. They also represent lost opportunity due to setting up of cells with suboptimal treatments.
Our approach allows working directly on historical business as usual data using a variant of uplift modeling. This approach effectively eliminates confounders and isolates the effect of treatments. A time series analysis is then applied, and FICO’s powerful scorecard technology is used to fit a model with node characteristics as predictors to easily estimate each node’s influence that it exerts on the rest of the network. This approach is not only cost effective, but highly scalable and can be applied easily to networks comprising millions of nodes and billions of edges. Further, use of historical data ensures that there are no missed opportunities.
Until recently, it has been very difficult to identify relevant social influencers. With the advent of social media, analytic scientists have gotten a powerful playground to understand peer influence and identify social influencers. This is made easy by the recent innovations in Big Data, which allow processing and analyzing of extremely large volumes of structured and unstructured data generated by social media platforms. These combined with FICO’s methodologies provide effective targeting at low cost to generated actionable insights and benefits to marketers in their peer-to-peer marketing campaigns.