In many industries there are inefficiencies solved by brute force because it is seemingly cheap to do so. In advertising, many publishers use the Spray and Pray technique, i.e, take a massive amount of advertising space and hope the right people will click on your advertisements. The reality is that conversion (from an exposed customer to a paying customer) can commonly be as low as 0.03%. This is a very low number indeed. Often, this number is not improved even after having rules like “show this advertisement to all women over 40”. Even though it may sound logical, it won’t improve much the result. The reality is usually much more complicated. In the example of the “women over 40”, maybe most women already bought the item or totally loyal to another brand. The utopian solution would be to make a rule for every user separately. For obvious reasons, that would not be feasible.
The cutting edge solution today is to use Machine Learning to analyze the vertical of each advertisement and match it to the most likely person to click on the ad. This matching is done automatically by machine learning algorithms. A machine learning algorithm creates a model from users and ads data which can be viewed as automatically constructing the best combination of implied rules. Based on the model, it reaches a conclusion whether to reveal the advertisement to the user or not.
The data to the algorithms, coherency and usefulness is what makes the algorithm bring out its full potential. Therefore, the nature of the data needs to be thought trough carefully. Then, the raw data needs to be normalized and processed before passed on to a machine learning system. At this stage, we will need to carefully choose the right algorithm for the job. Finally, it is very important to tie it all together to be a stable and maintainable solution in a production environment.
Many industries besides advertisement can benefit from leveraging machine learning techniques to increase efficiency and profitability. The trick is to know which data with the combination of the right algorithm will get you there.