The recently concluded Predictive Analytics World 2014 in Boston, Massachusetts, was chock full of insights from organizations who have successfully implemented analytics in a variety of settings. A few points stood out and I will attempt to capture them here:
1. Focus on ROI measures: This is spoken of very often, but frequently in an attempt to develop the “right” or “perfect” model, the focus on ROI sometimes begins to waver. Being driven by ROI implies understanding which variables are controllable by the business, which data observations are of real interest and sometimes making adjustments to accomplish that. This may mean considering variables that are otherwise not significant, or oversampling certain data observations and so forth. But a relentless focus on ROI will yield the desired results.
2. Eschew Complexity: Seek simpler models, fewer variables, and explanations that make sense. Given results, the human mind will find ways to explain it — so do not rely on interpretability as a defense of your models. But let the sheer simplicity of models tell their own story.
3. Ensure algorithmic Data Preparation: As all practitioners know, Data Cleansing and Preparation is 80% of the effort — but what is sometimes forgotten is that Data Preparation is not a one time effort, but is subject to the algorithms being considered. Understand not only the assumptions, but the limitations of the algorithms being considered — and do for the algorithms, what the algorithm cannot do for itself.
4. Consider Ensemble Techniques: Ensemble methods such as Random Forest, Gradient Boosting and others have proven repeatedly to provide stable and usable results. Master these techniques and more.
5. Simulations are often a good Communication technique — All practitioners understand that the ultimate success or failure of their efforts depends upon successful communication of their findings. Leverage simulation of your results, and the likelihood of success or failure of your model in real situations to help further communicate and define your findings.
And incidentally, if you wish to understand the analytics maturity of your organization, visit this link at the INFORMS website!
By: Sri Srikanth, Program Manager, Analytics, Digital Strategy & Enablement, Cisco
Originally published at http://blogs.cisco.com