As an e-commerce company, we aim to use computers to our advantage to serve our customer’s best interest. Because of this, we closely monitor all developments in the technology scene. Some of the most promising developments at the moment are found in the field of Artificial Intelligence. With “Machine Learning” at the peak of inflated expectations in 2016, it seems that everybody in the tech industry is excited about computers solving our problems and making predictions that are closer to the truth than humans could ever make.
We are no exception at Helloprint. The data in our organization is fuel for our growth. We monitor the activity in our stores in real-time so that we can act quickly upon problems. We use the data of our customer service department in order to make improve the user experience on our website. We use algorithms to provide our customers with the best deals available. In brief, we try to use the data we obtain to provide our customers with an even better user experience. The promise of Machine Learning, therefore, grabbed our attention and after some tests, we believe that we can use these techniques to help our customers even more. We must make the important side note, however, that it’s not a bed of roses. Computers are becoming “smart” and can solve problems, but they still need humans to prevent them from doing stupid things. Take for instance the case of Target: They use Machine Learning techniques to predict buying behavior and accidentally predicted (truthfully) that a teenager was pregnant in the US. Or Google Flu Trends which predicted more than double the amount of diseases than the Center for Disease and Control in America, causing an alarm of an outbreak that simply wasn’t there.
Nonetheless, if we learn from these lessons, Machine Learning techniques can live up to their potential. For instance, we have successfully completed the first tests for our recommender engine, where we can provide our customers with personalized recommendations based on the products that they have shown interest in. Moreover, we use Machine Learning to predict when customers will become inactive. This helps us to understand where we need to improve our services even more to satisfy our customers’ needs.
How can we prevent, however, that we will run into the same problems as Target and Google? Well, first of all, we need to avoid a black-box approach as much as possible. This implies that the technique needs to be understandable for everybody who works with it (at least on a high level). If we cannot achieve this, we are almost certain that people will use the method in a wrong way somewhere in the future. Secondly, we need feedback loops that are constantly monitored. The techniques need to be calibrated regularly to avoid overfitting. If the output does not make any sense or does not provide benefits for our customers; the model needs to change. Finally, we must not forget that sometimes less is more. Even though Machine Learning techniques are fancy and have a huge potential, we should only use them for jobs that we simply cannot do as humans. If a smart and easy solution solves the problem almost as well as the powerful Machine Learning technique, we might want to stick to it.
We believe that these three lessons are important to keep in mind. All in all, we want to use Machine Learning to improve our business. There are risks involved and only if one is aware of them, we can reach the full potential of the techniques. We hope that our prudence will eventually lead to more customer satisfaction and a healthier business.