10. Expect a new breed of model builder.
Azure Machine Learning (AML) is a disruptive technology that attracts all kinds of people who need model results. Unlike people investigating R or Python to expand their tool kits or solve new modeling problems, AML users know they need models even though they don’t know much about models and don’t have modeling tool kits.
Sound Bites from New Users:
- “Thank you! I’ve never built a model before!”
- “Thank you! …for making model building easy!”
AML is disruptive because: (a) it is inexpensive; (b) it is syndicatable (see point 6 below) and makes machine learning invisibly implementable*; (c) it is acceptable for myriad applications of machine learning; and (d) has an easy-to-use interface.
AML’s disruptive power attracts demand as users recognize they can reap machine learning benefits without investing a lot of cash into fixed investments such as: servers, appliances, support staff, and networking. Users can also skip the pomposity, self-entitlement, and laziness of today’s statistical software suppliers.
09. Ignore your statistical training.
Start your AML journey by exploiting toy problems and focusing on getting the feel of working in Azure Machine Learning Studio.
Memorizing formulas, reading esoteric material, and going over too much information too fast doesn’t help students develop a feel or intuition for data. Statistics taught in this way obliterates critical thinking by students while they are crunching numbers.
AML is a fresh start. Ignore any stats training you’ve had. Start by using AML to develop a feel for data, and intuitions about how models extrude data into new shapes and sizes.
08. Enjoy a new modeling paradigm.
Every discipline that builds quantitative models has its own tradition. Biostatisticians have a strong tradition. AIl researchers have another. The Chicago center of the marketing mix universe has another tradition. AML is 100% machine learning, and it uses the “train then test” modeling tradition of the machine learning tradition.
This may initially frustrate you if you come from another tradition. But let yourself enjoy the process of absorbing the machine learning tradition. It is valid even though it is different.
Every new tradition of modeling you learn, brings you closer to becoming a “rainbow unicorn” otherwise known as “data scientist.” Expanding your mind to hold multiple modeling traditions in your head at once should become your goal.
07. If you are from a statistical modeling tradition, relax and enjoy not having test statistics.
Azure Machine Learning does not give you the test statistics you are used to. For example, when doing regressions, you don’t get to see p-values because Azure Machine Learning is a tool, like all the other tools that you have collected on your modeling journey. What matters are the new capabilities the tool adds to your portfolio. If you are dependent on p-values, you probably don’t need AML to give you p-values.
AML gives you a new natural drag and drop modeling idiom as well as a massive expressive power in syndicatability. You can develop your model, deploy it in AML, and then use AML’s excellent Application Programming Interface (API) to connect machine learning directly to your users.
Simply install the AML add-in to Excel, then connect Excel to your AML model, and your users can apply AML to business problems via Excel, a tool they already know and love. Here is what the architecture looks like:
In a way, AML is the ultimate add-in for Excel. You get regression and random forests, Bayesian regression and many more time tested machine learning models.
AML will not replace your tool kit, but it will extend it in ways no statistical environment has yet done.
06. Syndicate a model as quickly as you can.
Round tripping is when you have an Excel model connected to AML, and you change data in Excel so it flows to AML in the cloud, gets crunched, and then you return predictions back to Excel. From Excel to AML and back to Excel = 1 round trip of data.
Syndicating from Excel to the AML API is the secret on-ramp to practical relevance in applying machine learning. You can take the operations Excel workbooks you rend what users are using now, add a sheet to the workbook, and implement machine learning invisibly. Your users will get the predictions they need without disturbance to their workflow. They don’t even need to see the AML worksheet since you can reference it from their existing rows and columns.
05. Cross validate. Cross validate. Cross validate.
Run your analysis in AML, then run it in R, Python, SAS, SPSS, SYSTAT. The process of replicating quantitative models always brings mistakes to light. And, as you model across environments, you develop an almost stereoscopic vision in your modeling. AML is a tool as are all other quantitative environments. But because no one tool does everything, do as much cross environment modeling as you can.
If you use multiple environments now (such as Excel, PowerPivot, PowerBI, R, SYSTAT, and SAS) you gain the ability to set problems up so they feel natural to end users, and so the problems and solutions give users expressive power in solving problems.
04. Just do it!
Running a single machine learning model is worth more than reading 100 blog posts on machine learning. Eight steps are involved in running your first AML model, and looking at output:
- Go to AML’s web site,
- Click “Get Started Now” and sign up to try AML for free
- Log in to AML and click on “Samples” and go to Sample 2:
- Then double-click Sample 2
- Read through the experiment and see if you can follow what is happening
Every box in the flow chart does one logical thing. Sample 2 is a great place to start because the boxes have good descriptions. Yes, they describe technical steps, but the entire experiment flows step by step. Building an experiment is similar to drawing in PowerPoint by inserting shapes, and then dragging lines between the shapes to connect them. That is how this model was constructed.
- Now, click the Save As button
- … and save the experiment as “Run it –Break it- Fix it
- Now click on “Experiments”at the upper left of the page and you’ll see your saved experiment:
- … now click “Run it –Break it –Fix it”
- “Run it –Break it –Fix it”will now fill your screen
… now click “Run”and watch as the experiment executes. You will see the experiment execute as green check marks appear (see red horizontal arrows above) and clock icons appear on steps waiting to execute (green vertical arrows above). It takes a minute. Be patient as the entire experiment runs.
- Explore the output of your experiment:
- Click on the bottom dot of boxes to pop open a dialog box. Then click visualize.
You have now run a machine learning model. And, you have learned the one basic skill: clicking on the dots of boxes necessary to explore the model and allow yourself to develop an intuition about what is happening in the experiment.Now, take some time and click around the experiment and let your mind absorb the Azure Machine Learning Studio’s way of organizing and processing data.
03. Exploit the Azure Machine Learning Gallery.
Take a model from the gallery, replicate the model (see point 05. above). See how genius data scientists at Microsoft and elsewhere actually do modeling in AML. Play, learn, repeat.
Try a model, break it, then fix it again. Gain a feel for how the model works in AML. Then try a model, add your own data, and see what comes out. Explore. Life is not about performance and perfection. It is about discovery. Discovery will grow your machine learning capabilities ten times faster than any other approach.
02. Sign out of Azure Machine Learning studio
Click on the user icon at upper right to open a dialog box allowing you to choose “Sign Out”
01. Close your browser. Then reopen it as “Sign In”again
If you lost the URL, it is: https://studio.azureml.net/
To log in again, go to https://studio.azureml.net/ and look in the upper right for “Sign In”. It may be a little tricky to find, but it is there.
You only need to create your account once. Do not sign up for another.