The Class of Learning Machines
Supervised Learning and BigML
“Machine Learning is like teenage sex: everyone talks about it, nobody really knows how to do it, everyone thinks everyone else is doing it, so everyone claims they are doing it.” Dan Ariely (who originally was referring to Big Data)
What is Machine Learning?
Machine learning relates to a branch of Artificial intelligence where the application learns through the data it processes. It is a tool that can get you from data to decisions, from information to insights.
In order to learn, a learning algorithm (set of rules) must be supplied with training data. The analytical model learns from patterns it identifies in that data. If you supply a model with a representative data set and the model generates certain predictions, the more data you supply the more your model will improve and learn. The model allows you to produce repeatable results from historical relationships and trends. It is only as good as the data you use to train your data model and will only work if the data you use is representative.
This quote from Pedro Domingues sums things up well:
“ Machine learning is not magic; it can’t get something from nothing. What it does is get more from less… Learners combine knowledge with data to grow programs “
Machine learning has been around for decades. So why all the machine buzzwords and bandwagons lately? The cost and speed of computing resources, data volumes, improvements in related algorithms and platform accessibility have all contributed.
Types of Machine Learning
There are in general 3 types of machine learning algorithms:
- Supervised learning where the algorithm is supplied labeled data with the correct output, from which the algorithm learns patterns. Best at modeling relationships and dependencies between data in order to predict future values such as true/false (classification) or some numeric variable (regression)
- Unsupervised learning where the algorithm is supplied unlabeled data and there is no guidance provided, and it would be difficult for a human to identify patterns and insights. Best at pattern detection and data grouping in order to provide insights
- Reinforcement learning whereby the algorithm learns from the environment in an iterative manner. There is a reward-feedback loop provided to the algorithm in order to reinforce learning. Best at taking action to increase reward and reduce risk, examples include self driving cars.
Customer Churn Prediction — Supervised Learning Classification
What if we could predict which customers are likely to churn and adjust our approach early enough? What if we could identify the customer value proposition that has the best outcome for our various segments? Predictive analytics can play a part in a retention strategy and is a great first step with machine learning before progressing to more ambitious challenges.
Some other interesting examples where machine learning is used include Little Ripper shark detection, the chihuhua or muffin resemblance which can be tackled via deep learning (a subset of ML) and determining voting preference. Another is Amazon Rekognition being used to identify persons of interest.
The Platform — BigML
In our example of a supervised learning classification challenge, we have used the BigML platform and a sample dataset adjusted to have some relevance to a financial advice practice. The dataset contains the historical results of who has churned.
The logic is based off of this comprehensive post, which predicts churn from a mobile Telco and utilises the Amazon Machine learning platform.
The Solution
We used BigML to answer the question of how likely a customer is to leave a business if we adjust certain factors such as service fee, number of days with business, whether our new advice model was applied etc. You can see an overview of the solution here.
The algorithm used is a decision tree, which identifies patterns in the data through if-then statements. The model provides a confidence rating for the prediction based on the number of instances of that particular if-then branch in the tree. Every tree is a story within the data that the model uses to predict the churn likelihood. The model also helps us identify which are the most important variables for our churn outcome.
This is only one example of what you can achieve on the BigML platform. It is also possible to use unsupervised learning techniques such as clustering and anomaly detection, amongst others.
What we Learned
- Take simple steps on accessible platforms before moving on to more complex business challenges
- Before getting value from machine learning the biggest challenge you are likely to have is identifying the right questions to ask.
- The next biggest challenge will be cleaning and normalizing your data. Feature engineering expertise will likely be required within your dataset
Machine learning can offer efficiencies in understanding our data. It cannot fully replace human perception and judgement. Correlation does not imply causation and therefore use data relationships as a guide only to adjust your approach.
The purpose here is to share simple practical solutions on accessible platforms that could be useful to test out your ideas. It’s not to delve into lots of technical detail and compare options.
Thanks to Lucy Adelaide for working on the illustrations above.
The views expressed in this article are those of mine alone. This is part 4 of a 6 part series exploring practical business applications of artificial intelligence.
