Social Network For Analytic Professionals
Interview with Dr.Eric Siegel, President of Prediction Impact, Inc.
Eric Siegel, Ph.D. is the president of Prediction Impact, Inc. and the program chair of Predictive Analytics World, the business-focused conference for predictive analytics professionals, managers and commercial practitioners, coming to San Francisco February 18-19, 2009. An expert in predictive analytics and data mining and a former computer science professor at Columbia University, Eric co-founded two software companies for customer profiling and data mining, and then started Prediction Impact in 2003, providing predictive analytics services and training to medium through Fortune 100 companies. Eric is the instructor of the acclaimed training program, Predictive Analytics for Business, Marketing and Web.
Which analytical fields are likely to experience growth, and why?
My work - consulting, training and the PAW conference - focuses entirely on predictive analytics, which is the application of predictive modeling to produce a predictive score for each customer or prospect. Each customer's predictive score informs what action to take with that customer - business intelligence just doesn't get more actionable than that!
Which methodologies might become obsolete, which ones are likely to entertain growth?
First I should say that, with predictive analytics, the modeling method itself is usually less of a factor for success than the precise way you use it for your business and data you give it - prepare thy data well!
But if you'd like to focus on the core method, the answer is, it depends. For many business applications, the most important aspect is the usability of the software, and the simplicity and interpretability of the predictive model - more complex models may be more fragile and less robust unless driven by a deep statistical experts.
But then on the other end of the spectrum, in situations where core analytical performance becomes the highest priority, the most promising area is "ensemble models", means to combine models and create an "uber-model" that is greater than the sum of its parts. In some cases this is a composite of many simple models, such as in the session by Dean Abbott at the PAW conference (http://www.predictiveanalyticsworld.com/agenda.php#nra). In other cases it is a matter of taking two very complex models and making an even better model, as in the case of the two Netflix Prize competitors who recently joined forces to create the leading competitor, "BellKor in BigChaos", who's also presenting at the PAW conference (http://www.predictiveanalyticsworld.com/agenda.php#advancedapproaches).
What do you recommend for students starting an analytical career or choosing a University curriculum?
Ultimately what speaks the loudest to your prospective employer will be relevant commercial experience. You need to have performed the same kind of analysis for the same kind of business case/problem. If necessary, do it once for free and a couple times for very little money to get started, because you need that in place. If you'd like to get a relevant degree beforehand, the specific area of study is really much less important. What you get in school is general analytical discipline, so pick the quantitative or scientific area of study you enjoy the most, or that you feel is most exciting, so you'll be fully engaged. In the end, "quantitative degree + relevant real-world experience" is the formula, where a higher degree-level helps, but the real-world experience trumps it.
What are the biggest successes of data mining and statistical sciences in the corporate world?
As a "recovering academic" who's been in the business world for several years, I'm now prone to ROI as the primary measure of success. With this in mind, success hinges on the business process that drives the analytics initiative. Tom Davenport had it right - executive buy-in and an organization-wide adoption of analytics makes all the difference, whether you're performing customer retention with attrition modeling, response modeling for direct marketing or producing product recommendations. With that in mind let's turn to your next question:
What are the best practices for analytic professionals?
The integration of predictive scores to affect decision automation or decision support for your business - whether targeting marketing offers and sales resources, or affecting the product recommendations issued on your website - must be planned out careful in advance, in all its gory details. You need to start by completely defining this end-goal, intended deployment of a predictive model. This in turn determines the data preparation process.
In a nutshell, this best practice organizational process is described by the often-cited CRISP-DM. But that standard doesn't solve the dilemma all by itself. There are two problems. One is, CRISP-DM can be somewhat abstract or academic to a first-time reader - try reading this short, digestible intro first: http://www.predictionimpact.com/customer-prediction.html
The other problem is that even the most senior expert struggles to truly accept how involved and time-consuming the project planning and data preparation phases are. I speak from personal experience -one kind of lives in denial because you just don't want to believe there will be that many gotchas, roadblocks and bottlenecks as you collect, transform and debug the data. It's almost always worth the hassle, but it's got to be planned out right. I think the best antidote is to adopt the "CRISP-DM" mantra. Wake up every morning and chant "CRISP CRISP CRISP" 42 times!
How and why did you become an analytic professional?
My graduate research was in core predictive modeling as an abstraction, whether to solve business, science, engineering or linguistic problems. In the academic world, this area is called "machine learning". The idea of a computer learning from experience by somehow deriving patterns that hold in general - within the data at hand and beyond - has always fascinated me. So, I think it is the coolest of the cool. What more profound ambition could we have for a machine than to "learn" in this sense? Turning to the business application of this approach, it does indeed turn out that there's a real value to be gained by predictively scoring customers - so, in that sense, it's not only cool, but also exciting.
Which analytical fields are likely to experience growth, and why?
My work - consulting, training and the PAW conference - focuses entirely on predictive analytics, which is the application of predictive modeling to produce a predictive score for each customer or prospect. Each customer's predictive score informs what action to take with that customer - business intelligence just doesn't get more actionable than that!
Which methodologies might become obsolete, which ones are likely to entertain growth?
First I should say that, with predictive analytics, the modeling method itself is usually less of a factor for success than the precise way you use it for your business and data you give it - prepare thy data well!
But if you'd like to focus on the core method, the answer is, it depends. For many business applications, the most important aspect is the usability of the software, and the simplicity and interpretability of the predictive model - more complex models may be more fragile and less robust unless driven by a deep statistical experts.
But then on the other end of the spectrum, in situations where core analytical performance becomes the highest priority, the most promising area is "ensemble models", means to combine models and create an "uber-model" that is greater than the sum of its parts. In some cases this is a composite of many simple models, such as in the session by Dean Abbott at the PAW conference (http://www.predictiveanalyticsworld.com/agenda.php#nra). In other cases it is a matter of taking two very complex models and making an even better model, as in the case of the two Netflix Prize competitors who recently joined forces to create the leading competitor, "BellKor in BigChaos", who's also presenting at the PAW conference (http://www.predictiveanalyticsworld.com/agenda.php#advancedapproaches).
What do you recommend for students starting an analytical career or choosing a University curriculum?
Ultimately what speaks the loudest to your prospective employer will be relevant commercial experience. You need to have performed the same kind of analysis for the same kind of business case/problem. If necessary, do it once for free and a couple times for very little money to get started, because you need that in place. If you'd like to get a relevant degree beforehand, the specific area of study is really much less important. What you get in school is general analytical discipline, so pick the quantitative or scientific area of study you enjoy the most, or that you feel is most exciting, so you'll be fully engaged. In the end, "quantitative degree + relevant real-world experience" is the formula, where a higher degree-level helps, but the real-world experience trumps it.
What are the biggest successes of data mining and statistical sciences in the corporate world?
As a "recovering academic" who's been in the business world for several years, I'm now prone to ROI as the primary measure of success. With this in mind, success hinges on the business process that drives the analytics initiative. Tom Davenport had it right - executive buy-in and an organization-wide adoption of analytics makes all the difference, whether you're performing customer retention with attrition modeling, response modeling for direct marketing or producing product recommendations. With that in mind let's turn to your next question:
What are the best practices for analytic professionals?
The integration of predictive scores to affect decision automation or decision support for your business - whether targeting marketing offers and sales resources, or affecting the product recommendations issued on your website - must be planned out careful in advance, in all its gory details. You need to start by completely defining this end-goal, intended deployment of a predictive model. This in turn determines the data preparation process.
In a nutshell, this best practice organizational process is described by the often-cited CRISP-DM. But that standard doesn't solve the dilemma all by itself. There are two problems. One is, CRISP-DM can be somewhat abstract or academic to a first-time reader - try reading this short, digestible intro first: http://www.predictionimpact.com/customer-prediction.html
The other problem is that even the most senior expert struggles to truly accept how involved and time-consuming the project planning and data preparation phases are. I speak from personal experience -one kind of lives in denial because you just don't want to believe there will be that many gotchas, roadblocks and bottlenecks as you collect, transform and debug the data. It's almost always worth the hassle, but it's got to be planned out right. I think the best antidote is to adopt the "CRISP-DM" mantra. Wake up every morning and chant "CRISP CRISP CRISP" 42 times!
How and why did you become an analytic professional?
My graduate research was in core predictive modeling as an abstraction, whether to solve business, science, engineering or linguistic problems. In the academic world, this area is called "machine learning". The idea of a computer learning from experience by somehow deriving patterns that hold in general - within the data at hand and beyond - has always fascinated me. So, I think it is the coolest of the cool. What more profound ambition could we have for a machine than to "learn" in this sense? Turning to the business application of this approach, it does indeed turn out that there's a real value to be gained by predictively scoring customers - so, in that sense, it's not only cool, but also exciting.
Tags: eric iegel, prediction impact, predictive modeling, san francisco
Featured
Advertisement
Analytic Jobs
Events
© 2010 Created by Vincent Granville
