Machine learning uses algorithms to analyse data.
The purpose being to identify patterns within the data. Once identified the patterns can be used to make predictions about specific cases as they arise.
Machine learning then is analogous to human behaviour. Humans look at what goes on around themselves and then draw conclusions as to how the world works. They can then apply what they have learnt to new situations as they arise.
One of the most widely used applications of machine learning is predictive modelling or predictive analytics.
The tools derived from machine learning can be used to determine something that is not already known based on the data already available.
Often this relates to the past behaviour of an individual and subsequently what they went on to do. It is then possible to make a future prediction of a given individual’s behaviour based on their current activity.
In marketing predictive analytics can be used to identify the purchasing history of a customer. A business can then utilise this information to predict what they will purchase next. This allows the business to target the customer with product specific marketing offers.
This is just the beginning. Already there are advanced predictive modelling tools, such as Optimove, which have the capability to access business data and generate models that display Customer Lifetime Value. From here businesses can make informed choices about future marketing campaigns and product development.
Another way to understand machine learning is to look at it in terms of reducing uncertainty about the future. Whilst machine learning cannot offer certainties relating to a given outcome, it can provide valuable insight.
A grocery store may know that when one of their customers goes shopping that they purchase bread, wine and chicken. Machine learning can calculate that there is an 80% probability that they will be bread as opposed to a 15% and 5% chance that they will buy chicken and wine respectively. It makes sense therefore to offer that customer a bread offer to encourage them to make their next purchase in that store.
Thus a predictive model identifies the patterns uncovered by the analytical process. Once created the model can be used to generate new predictions.
An organisation can then use the model’s predictions to decide how to treat individuals.
Machine learning is the process, predictive modelling is the end product of that process.
While there are many types of predictive modelling and thousands of machine learning techniques a model’s predictions are nearly always represented as a score. The higher the score the more likely the individual is to behave in the way the model predicts.
Applications of Machine Learning
1 – Credit Scoring
A well known application of machine learning is credit scoring.
When an individual applies for a loan, a credit card or a mortgage the bank or advisor will ask them a series of questions based on their lifestyle. This information is then cross referenced with a credit report that details the individual’s previous financial history. This information is then fed into a predictive model to generate a credit score.
In the USA a FICO score of over 750 is a prediction that an individual is credit worthy, that is likely to be able to repay a loan. Whereas a score of less than 500 indicates that an individual is much less credit worthy.
2 – Marketing
A second well known application of machine learning is target marketing.
Having information relating to customers, such as age, gender, income, web-browsing, purchase history, location, etc can help to predict if an individual has an interest in a certain product or not. Consequently a marketing department can decide which product or offer to target each customer with based on the prediction.
In marketing predictive modelling can be used to tailor a personalised pricing strategy to each individual’s circumstances.
3 – Health Care
One of the most beneficial uses of machine learning is in preventative health care.
Traditionally people look for medical assistance when they become ill. Doctors then respond by treating individual cases as they see fit, which can be time consuming and costly.
Machine learning can evaluate an individual’s medical records and predict the likelihood of them developing specific conditions such as heart disease or diabetes many years in advance. An individual with a high predictive score is the most likely to contract a particular disease and so can be contacted with a view of initiating a preventative action.
4 – Social Media
A final example is how machine learning is used to determine what types of content to recommend to individuals on news and social networks. Media providers analyse what articles an individual has read in the past and what comments they have made in response to those articles. The provider can then conclude what is appropriate content to promote.
Other common applications of machine learning include the algorithms that match people on dating sites, the technology used by payment processors to detect credit card fraud and systems used to identify terror suspects.
The graph from Venture Scanning below illustrates the breakdown of machine learning usage in relation to the amount of money being invested in machine learning.
As you can see the Applications category is leading with over $2 billion market share. Thats three times greater than the total funding of the next category: Natural Learning Processing.
The term artificial intelligence is a broad one that can be used to apply to machine learning and predictive modelling. However, whilst artificial intelligence does rely heavily on machine learning it is incorrect to say that they are on and the same.
For the purposes of this article I will outline artificial intelligence as the replication of human analytical and decision making capabilities.
A good AI application is one that can perform an operation as good as or better than a human. This could include an assessment of an individual’s credit worthiness, the ability to identify people from their Facebook profile pictures, the ability to beat the best chess player or being able to spot early signs of cancer on a medical scan.
In short, i think it’s fair to say that Artificial Intelligence can be understood to be highly sophisticated applications of machine learning.
The core components of artificial intelligence are :
- Data Input: This can include sensory inputs from microphones or cameras, processed data when an individual fills out a form online, an individual’s purchasing history or an individual’s credit check.
- Data (Pre-Processing): Raw data needs to be pre processed into a suitable format before it can be used for machine learning
- Predictive Models: These are generated from the machine learning process using large volumes of historic data. Pre processed data is fed into the models with the aim of generating future predictions.
- Decision Rules (Sets): These are used alongside the data and predictive models to decide what course of action to take. These maybe defined automatically via the machine learning process but also by human experts or business users.
- Response / Output: Once a decision is made an action must be taken. For example, once a bank has ascertained that an individual is credit worthy a loan can be issued.
The combination of all of these components provide us with the artificial intelligence. The sheer complexity of some of the underlining algorithms is what makes artificial intelligence so impressive. Couple that with a slick user interface that presents the findings in a clear and actionable way and you have a serious AI!
Artificial Intelligence and Marketing
Consider the example of a drinks company that is looking to gather information about individuals from social networks. It then wants to feed the information into a predictive model to ascertain how many of those individuals would be interested in purchasing a certain brand of whiskey.
The predictive model would then apply a number of rules which may include:
- If the likelihood of the individuals buying the whiskey is 90% do nothing – they will most likely buy it anyway
- If the likelihood is between 1% and 90%, send them a $5 discount coupon – this will influence their behaviour and increase the chances of them purchasing
- If the likelihood is 1% do nothing – they are unlikely to buy the whiskey no matter what is offered to them
These rules would define a cost benefit analysis of the optimal level for triggering marketing activity.
A further set of rules would come into play based on regulations and a business code of ethics:
- Alcohol, in this case whiskey, should never be sold to children, no matter how likely they are to buy it.
- Do not sell whiskey to individuals that have a history of alcohol related problems.
Both of these two groups are likely to contain a high number of individuals who would probably buy the whiskey. Whilst it may be profitable to sell to these two groups it is likely to attract much unwanted negative publicity.
These two rules highlight exactly why it is valuable to have human input in the decision making process of an AI, especially when called upon to make risky or controversial conclusions about people.
The next step is to link the AI to the drinks company’s marketing channels. The AI may decipher what action is to be taken but still a human is needed to design the marketing content, format the wording and create a communication that is appealing and compliant with the relevant marketing regulations.
Machine Learning and Artificial Intelligence Today
The rise in popularity of smartphones, tablets and the internet has provided more information about users than ever before. Along with this computer storage capability has increased to manage this ever growing supply of data.
As a result the computer algorithms have grown more complex and the growth of machine learning has seen an exponential increase.
Nowadays there are a whole host of machine learning tools. Programming languages such as Python and R are open source and made available by Google and Microsoft for developers to use free of charge.
Not only that, but there are competition sites such as Kaggle. Here organisations can present their data whilst amateur and professional teams compete to build the most predictive models based on the data.
Take a look at the Video below to see how Kaggle works:
Consequently AI systems are able to interact with people in ever more human ways.
Often interfaces between computers and humans are powered by predictive models.
Returning to the whiskey example from before, machine learning could be used to build a chatbot that is able to communicate with users on social media to ask them questions about whiskey.
The data collected can then supplement other previously gathered data. The whiskey application can then use a predictive model to determine which users could be offered a promotional discount and send it back to them via the chatbot on social media.
The result of all of this is that machine learning is now no longer the domain of large businesses and academic institutions. Anyone with a laptop and access to the internet can start building predictive models and develop this into an interface to implement those models as an AI application.