Moderator: One to two-day training is a lot better than six-week training. Talk about models and scoring models and managing different models over time. A lot of these data mining applications can be set up such that the scoring is done automatically.
In other words when you apply this algorithm for “next best offer” that sales increase 2.1%, for example, whereas that other model, sales actually decreased to 0.2% or 0.7% or something like that. I am sure some of that stuff can be automated, but what are some of the other best practices around scoring models?
Flaherty: Definitely, I think for scoring of models, we have seen lot of automation and handholding with IT. How can I automate that critical step and avoid any kind of manual kind of intervention?
You are try to reduce the revalidation efforts. I think one critical best practice in terms of database scoring is for where you are trying to take the scoring algorithm and calculate the score of new set of data. The other best practice is how can continuously monitor the performance of your predictive model.
How Can I Monitor the Performance of a Lead Scoring Model?
If my model is not performing well, how will I know? How can I monitor that performance and bring that model back into the production cycle. And as part of that issue, then there is the question of how do I do that. Some of that performance monitoring, which include charts or reports or alerts, is being integrated with scoring in a management framework of analytical models. So that’s an important consideration. How can I avoid risk in terms of model degradation and then continue to guarantee the usefulness and accuracy of the model? Combining the scoring techniques along with the performance monitoring is an important best practice.
Moderator: What about best practices for data mining in terms of building flexibility into your campaigns and into your program for data mining? It’s just amazing how fast you can learn things and how fast you can retool what you are doing. You can refocus your efforts. You get instantaneous feedback. It’s not like the old days when you would send out your direct mail piece and six weeks later maybe you have got enough data to analyze a campaign. What is some advice you can give to companies to help them build flexibility into their programs?
Flaherty: It’s a good question I think this is a really exciting topic. The most important thing has been the deployment of a model and getting the right apparatus in place to that deployment. So the way that I look at it, first there is the real-time aspect of your predictive models running there, providing value to your business. They are allowing you to monitor how things are going at all times.
At the same time you need to provide flexibility. One of the key changes in the market recently in terms of best practices is to look at your data sources that are now out there. Ask what insight they can give you for your business. And so you have a triangulation going on. You have got the social media sphere. You know that’s a different kind of data. You don’t know necessarily who is seeing those comments and tweets, but you know the types of things that are being said and the trends.
“Flexible product with great training and support. The product has been very useful for quickly creating dashboards and data views. Support and training has always been available to us and quick to respond.
- George R, Information Technology Specialist at Sonepar USA
Number one is the real-time monitoring aspect of data mining models’ performance. Next is having the tool setup to automate alternating among models. Then comes the idea to be open and consider all types of data sources that can add value to your analysis. Don’t ignore the huge changes going on in how people buy things today and how they communicate with one another.