Blogs

Social Media Analytics, ROI and Other Non Sequiturs

Solution CTO, IBM

Okay, I’ve bitten my tongue for as long as possible, and I need to get this off my chest. Most, and by most I mean at least 85% of the social media big data initiatives I see are ultimately going to under-deliver their expected ROI or eventually be discarded as not useful enough. I was reminded of this recently via a conversation on LinkedIn where someone asked a very ill-defined high-level question and 10+ people jumped in with product recommendations.

Wrong. Wrong. Wrong. Throwing tools at the problem without knowing what the problem requires to be solved is absolutely a worst practice.

Before doing anything, you need to explain and document the problem in as much detail as possible. What are you are trying to solve? Do you want application- or web property-specific trending? Topic aggregation or topic evolution? Behavioral or interaction insight? Is it important to get down to a personal level, or is keeping things at a crowd level okay? Do you need to build influence models? How quickly do you need to be able to ingest inputs? How long do you want to persist the data? Do you need to resolve entities against an internal database?

Doing anything before you have a roadmap that addresses the requirements that will lead to a sustainable (that is, positive ROI) use case is a mistake. Let me say it again - throwing products or tools at this without understanding the questions above is a worst practice, and one of the reasons so many of these projects fail to deliver sustainable value. The only exception I can think of is if you don't have any visibility into what people are saying about your company or product. In that case harvesting company and product mentions is probably fine, but don't expect it to do anything except give you some visibility into what people are saying about you.

Knowing what someone said is far less interesting than why they said it and understanding what you can do about it. Doing that requires blending internal and exernal information. Are people responding to something that you have done, such as revised policies, products, pricing, hours, or one of the thousands of other things that you do in your day-to-day business? If you can't relate and build predictive models of how these changes impact and drive people’s interactions with you in the social space, then you're just as likely to get it wrong as right. More importantly I think it's inevitable that you're going to have to link the two, so why take a path that will prevent you from doing exactly that?

You need to build models that understand how what you do inside the building, impacts what people are saying and how they're responding outside, You also need to build models that adapt what you do inside the building, based on how people are behaving outside of the building, and then assess the effectiveness of your response.

Keep in mind there are ultimately two tasks here.

  1. Build the models, and
  2. Score the models extremely quickly.

This is one of the reasons why we promote models directly from InfoSphere BigInsights into InfoSphere Streams for real-time scoring, so you can act on patterns as they are happening rather than hours, days or weeks after the fact. Having a model but not being able to exercise it quickly enough to drive the right outcome pretty much defeats the purpose of the effort.