Articles - How does Microsoft’s new machine learning platform stack up?

 by David Hughes, associate at Barnett Waddingham 

 I’ve been looking for a slick solution to help me, my team and most importantly my clients. I want a solution that will;

 Before moving on, I want to just give a quick overview of what predictive modelling is and more what machine learning is.

  

 What is predictive analytics?

 The science of building models which can help predict outcomes is called predictive analytics. Much of the heavy mathematics has been around for years. Recent technology has allowed the proliferation of advanced methods which previously would have been too demanding for most PCs.

  

 Machine learning is one of the advanced disciplines within predictive analytics. It changes the way we can build models, using the power of modern computers to effectively learn from data.

  

 Predictive analytics is used everywhere. Spam filters, recommendation algorithms and retail customer analysis. If your business isn’t using predictive analytics to improve key business areas, then you may be falling away from the pack.

  

 Can Azure Machine Learning build predictive models quickly, easily and robustly?

 In short, yes it can.

  

 Uploading data is easy and straight forward. Once the data is in the environment, you can pick it up, connect with other data and do everything that you would expect to be able to do. I did struggle to connect to a database, but I’m sure with a bit more effort, and time, I’ll get it working from one of the Azure repositories.

  

 The graphical interface is intuitive and very easy to use. A simple drag and drop component approach enables one to build a model from data down to validation and testing.

  

 The component bar is on the left, and gives you quick access to all the tools you need. It’s very well organised, and comes with a quick search box at the top. Once I had completed my first model, I felt comfortable navigating and finding components.
 I created my first model using a sample dataset, there are numerous example sets included. They are nice and easy to manage.

  

 For my second model, I used a set of motor insurance telemetry data. Approximately 70,000 rows of data. I had already analysed this in the statistical package, R, and produced a couple of possible models.

  

 The R modelling took about a day, including using the business intelligence tool Qlikview to visualise some of the multi-variate relationships.

  

 The Azure model took me about 30 minutes to build and test.

  

 Can Azure Machine Learning create a seamless transition for Data to modelling to deployment?

 A key selling attribute of the Azure ML platform is ability to rapidly deploy a model. Once you’ve built it and tested it, you can publish it and create an API to have a cloud based scoring routine.
 Microsoft is clearly putting huge bets on the future of the Azure platform. I’m looking at how all the offerings can interact together.

  

 And part of that is the data interconnectivity between the Azure data platform and the ML side. I’ve yet to test this fully, but all the ingredients are together.

  

 I expect this to become a powerful, integrated environment.

  

 Does the Azure Machine Learning platform enable a wider skillset to use the modelling?

 The combination of the easy to use components, nice interface and logical modelling layout gives me a great sense that an individual with little technical knowledge of machine learning can build a model.
 But…
 I still feel that one needs to have experience and understanding of the modelling process.

  

 The Azure platform makes modelling easy. And has potential to open up the powerful tools of machine learning to less ‘data science’ trained people. This is good, given the shortage of data scientists, and high salaries the good ones command.
 Yet, there does need to be some base training and support around building models, and how predictive analytics can really impact a process.

  

 Reading this post, you expect me to encourage you to go out and get Azure ML. And yes, I potentially do encourage you to do so, however My one issue with it is that I can’t see the parameterisation of the models.And that is a major issue for me. It’s too black box. If I build a simple tree in R, I can get a graphical representation of it. Which is critical to enable me to discuss the model with my clients.

  

 Does Azure ML need me to change the way I engage with clients and the end users of my models?
 Perhaps.
 When I raised this with the team at Azure, I got a quick response suggesting there is some functionality, and further releases will see this increase. The speed of the response made me wonder if MS had deployed a predictive test responder to my question.

  

 Will I be getting this for my team?

 I will be getting licenses for my team. But it won’t make my use of R or Python redundant. It is a powerful tool to add to my quite large list of tools in the data mining toolbox.