One of my passions in life is predictive analytics. Countless blogs and presentations have been inspired by my endless appetite to evaluate a myriad of predictive tools. A few weeks ago I noticed a peer at Microsoft posted a video on using Azure ML with Power BI. In his video, David Gollob showed visualizing Azure ML scored models from Azure SQL Database. Fantastic! I didn’t know that you could output Azure ML scored model results to Azure SQL Database. I could not help myself, I had to immediately give Azure ML another whirl.
The last time I played with Azure ML was a few months ago. At that time, I wanted to check out the newly added R functionality. It was nice but I struggled trying to consume the scored predictive model web service output. It was not nearly as easy as it is right now. Better yet, the newer Azure SQL Database connection in Power BI is a true “direct connect” meaning that I can simply operationalize the predictive model results into a business process with an intelligent dashboard.
Exploring Bike Buyers
Predictive analytics is not new to Microsoft. I was playing with the SQL Server 2000 Analysis Services Poisonous Mushroom Decision Trees demo data set a loooooong time ago. My favorite predictive analytics data set that I have used in most of my predictive tool evaluations from Weka, Rapid Miner, R, Alpine Data Labs, Alteryx, Big ML to TIBCO Spotfire, Tableau, Pedixion and Frontline Systems is Adventure Works Bike Buyers. I still like that one because it is small, trivial with regards to complexity, anyone can download it from CodePlex and I already know what the correct results should look like.
Taking the Adventure Works Bike Buyers csv flat file data set, I uploaded it to Azure ML Studio, split the data set into training and test groups, hooked up a decision tree algorithm, a scoring task, an evaluation task and a data output destination to my Azure SQL Database. After previewing the initial scored model results, ROC, lift and accuracy, I noted that they were indeed consistent with results found using other tools when I played with this same data set.
Now that I have a scored predictive model data set, I can merely connect to Azure SQL Database from Power BI Desktop or Power BI in the web browser to explore it. In the real world, I often will merge the scored data set with a data set that contains known results and use a data discovery tool to fine tune my predictive model design by looking at where the errors are happening. I find visually slicing and dicing predictive model errors is quite revealing. It helps me experiment and better prepare the data being predicted for optimal model results. The data prep aspect of data mining is an art and a science. Honestly, it is a critical success factor for achieving predictive model accuracy. Visualizing the data being predicted speeds up learning what works and what does not when experimenting in an iterative predictive data prep process.
Putting it All Together
Here is another little tip for you. I typically will add a model diagram and model definition along with the exploratory visuals. I do this for non-data science audiences to add more understandable context in the event they want to explore the results on their own. Since decision trees are not available in most data discovery tools, I will often take a screen shot and add the image to my report. With the recent Power BI D3.js visualization framework, I may try to add an interactive decision tree soon. Here is a peek at what my predictive model report looks like in the latest release of Power BI Desktop.
How to Get Started
If you want to try this yourself, you can do so for free. Sign up for a free Azure ML Studio account. If you already know a little about data mining or predictive analytics, the drag and drop Azure ML user interface is pretty easy to learn. At least, there was minimal learning curve for me transitioning from other predictive analytics tools. There are also free Azure ML videos, tutorials and sample templates that can be helpful to reverse engineer. To visualize the Azure ML scored model output, download a free Power BI Desktop app or sign up the free web version at PowerBI.com. That is all there is to it. Simple, free and powerful. If you enjoy data, I will confess that this specific technology combination is fun and addictive too.