Assuming you are data rich but information poor, machine learning could be a good investment, but not all machine learning (ML) dollars are equal.
Predictive model testing can calculate how much ‘lift,’ i.e., the gross benefit you can expect from following the model’s advice. Add in the costs of following the model’s advice, which may have some interesting steps and curves and not be the intercept + slope we all wish for, and you will have a forecasted net profit.
But how do you calculate the ROI of being allowed to create a model to test when you don’t even know what lift you might achieve? How do you convince someone to let you find out?
Just let me build it and we will have an embarrassment of success … maybe
Assuming you prosper by doing your core activity better than the competition, then it really depends. Edwin Chen’s blog provides an example. Real humans scored Amazon, B&N and Google Play recommendation engines for relevance. Three leading organizations presumably applying premium intellect to a core capability. Yet Edwin’s findings are Amazon scored 42% very good (recommendations), 18% very bad. B&N has a similar score, while Google Play scored 17% very good and 51% very bad.
Will it cost a lot of money to find out if machine learning will make my organization rich?
You will need:
- Machine learning software – Your software investment can vary wildly from open source (but read the box carefully as some assembly is required), to ‘our algorithms are 99% confident you can’t afford it, so don’t even inquire.’ It isn’t necessarily the largest expense, but in large organizations just provisioning it can be a barrier to getting started.
- Talent to select and train the predictive models – In the kdnuggets salary survey for 2014, the US/CA data science salary averages $128k USD, but appreciate they are social animals and may not thrive if alone for too long.
- Someone who can convince everyone else to actually let the model act upon the things your organization holds most precious – That is you, convincing your operations leadership, your risk team, your lawyers and your HR leadership to let the model do things to the employees, the customers, the assets and the facilities. If it isn’t allowed to do things, then nothing changes and your ML dollars are utterly wasted. No surprise, this is the hardest part of an ML project. The challenge is even greater when your model is generating predictions that must be acted on without any human approval steps as found in certain Machine-to-Machine use cases.
A competing proposition for locating software and talent
Offer a prize in an open competition. For those that have an interesting problem, you may lure lots of data scientists to unleash their best voodoo by offering a prize of $3000, but the winner expects their work to be put into orbit. A $40k USD prize for something more down to earth and $3MM USD for something that will take them years to figure out.
Or, do it yourself for zero down on software and hardware
Microsoft was an early participant in making machine learning accessible to regular business people who have a basic grasp of statistics. Excel 2013 has a data mining add-in and data mining has been in SQL Server since 2005.
More recently the Azure Machine Learning preview brings additional algorithms all for zero down. Your machine learning experiments are entirely a pay-as-you-learn expense and that pay-per-use approach still applies when you are ready to put it into production. My next blog post will explain how elastic this cost model is.
In Azure ML your models, and even model ensembles, are created using a browser presenting a visual drag-and-drop experience. You don’t need to be a coder or a data scientist to create models to transform your organization from being information poor into insight rich. Once you have convinced everyone to hand over control to the model, you can leave the lab and put it into practice. If embedded into your normal operations ML can immediately and automatically initiate action to create success.
The combination of ML predictions being an input into a decision management system that initiates or advances workflows in response to events is the essence of Digital Business . This is why in a Digital business context, ML can make a big impact because the insight captured during the modeling process is now a force-multiplier. When ML is combined with the other elements of Digital Business, like Intelligent Systems Service, you have your top talent, in their finest hour, using the freshest data on every decision, every time.