November 28, 2016 at 11:58 AM
Machine Learning is by no means a new thing. Back in 1959, Arthur Samuel’s self-training checkers algorithm had already reached “amateur status” – no mean feat for that period in time. This article is intended to shed some light on the two different types of Machine Learning that one can encounter, which may be useful if you are thinking of entering into this space and are unsure as to which avenue is appropriate for your business.
The two types of Machine Learning to consider
First of all, the machine learning field can be separated into two types. They go by different names, but are widely known as 1) static or batch and 2) dynamic/incremental or self-training. The two approaches differ in some key areas which we will describe in this article.
Machine Learning Type 1: Batch Models
Batch Models are models that are retrained at specific points in time and are generally retrained manually and ‘offline.’ Since the development of the models is manual in nature, they take longer to build.
Some examples of such models are churn or attrition models that identify with some degree of certainty the likelihood of someone leaving you for one of your competitors. The business application will determine how much time is spent on developing the models, as well as the expected life of the model.
For example, within the credit risk space, scorecards for predicting future payments need to perform well relatively far into the future and therefore much time is spent ensuring that the characteristics and associated patterns are stable and trustworthy. Depending on various factors, these models can be created in hours versus months (seriously!).
Machine Learning Type 2: Incremental Models
Incremental models on the other hand are ‘online’ models that retrain themselves with little or no manual intervention. Examples of such models sit within the call centre environment, where the underlying data changes on a regular basis and new patterns emerge and disappear at a more regular rate.
Setting incremental models up optimally involves adjusting how much influence recent data has on the parameters of the retrained model. Environments where trends in the underlying data may change frequently require models that place strong significance on new data, almost forgetting trends that have been observed further in the past.
On the other hand, environments where trends in underlying data stay relatively consistent will see models performing well over time by only slightly adjusting parameters based on new data. A good example where recent data is more relevant than data from 12 months ago is in the call centre space where, for example, a “Right Time to Call” model will feed off dynamically changing dialler data.
It is the setting up of the automated data feedback loop for retraining purposes that makes incremental Machine Learning models more complex, as well as the automation of the monitoring systems that are required to ensure that new models are robust enough before replacing the old ones. However, once you have this system set up, pointing to other areas of the business becomes much easier.
Choosing the right type of Machine Learning for your business
For some businesses, the idea of an automatically (even semi-automatically) retraining system in some areas of the business is appealing – again for example, in call centres to identify which agent should speak to which customer, or which phone number to call at which time of the day. These models can be developed fairly quickly, and automation of the retraining process can be applied with some comfort in the knowledge that an inherent weakness in the model will not lead to a significant cost to the business in the event of an under-performing model.
However, for some other applications with a higher consequence of failure, this approach is not wise. For example, a model that predicts which loan applications should be accepted and which should be declined, or what credit limits should be allocated has a higher consequence of failure. Getting these decisions wrong can cost businesses millions in the long run.
Read more about the different levels of consequences of failure in our blog post Making the move from Predictive Modelling to Machine Learning.
For these applications, greater effort in developing predictive models needs to be applied using stable features and only trusting consistent patterns that will lead to a more stable and trustworthy model. If you have experience in developing these static models, then you have the fundamentals required to move into the more dynamic self-training environment and to apply the benefits of Machine Learning to other areas of your business.
Not sure which type of Machine Learning is right for you?
For 18 years now, we have been developing some pretty effective and high performing propensity models for the financial services industry, covering retailers, insurance and telcos, as well as across diverse applications – such as predicting an individual’s propensity to churn, to become a new customer, to miss a payment, or to take up an additional product.
Image credit: iStock