How to train your inbox using Microsoft Azure Machine Learning (part one)

What is machine learning and how might it be useful?

Microsoft recently released Microsoft Azure Machine Learning, a suite of offerings that enable customers to easily create, test, implement and manage predictive analytics solutions in the cloud.

Predictive analytics makes use of patterns found in historical and transactional data to make future predictions. Credit scoring is one example, where financial services institutions use factors such as credit history and socio-economic profile to assess an individual’s ability to repay. Insurers also use predictive analytics to assess insurance risk.

Machine learning software is not new, but has been difficult and expensive to implement, requiring specialist developers or data scientists. By moving predictive analytics to the cloud and enabling developers to embed predictive analytics into their applications, Microsoft Azure Machine Learning changes that. I was keen to give it a go…

Taking Azure Machine Learning for a test-drive

Microsoft have done a good job providing product walkthroughs, so I started with their simple predictive analytics experiment.

This walkthrough covers the steps required to predict automobile prices based on various factors such as make, body style, wheel base, engine size, horsepower, peak rpm and highway mpg.

Experiments consist of a number of datasets and modules, which can be dragged onto an experiment designer:

‍

Once complete the automobile price prediction experiment looked like this:

‍

‍

Having successfully completed the walkthrough and publishing as a web service I was suitably pleased with the results.

This ability to create experiments, train models using historical data and then test new results via a web service creates interesting opportunities for building business apps that make use of trained data.

Just one example: in an ecommerce application you could create a fraud-monitoring step in the order placement and fulfillment pipeline. Based on historical data factors such as product type, price, new customer, delivery address an alerting system could be used to place orders on hold for a team to authorize or halt potentially fraudulent orders when they have a certain set of characteristics.

In my next two posts, I’ll show you my first experiments with machine learning, applied to a common source of pain in the workplace: the inbox.