Boosting Customer Experience with Neural Networks

Rajeeve Kaul, Corp Vice President, McDonald's [NYSE: MCD]As interest in AI and neural networks has grown, there is a growing body of evidence that these concepts work better in some business situations. However, for business users, there is a lack of good clarity on what those situations are where one can relatively, reliably expect to see benefits. Exemplars so far have touted the benefits of these new methods on handwriting recognition, natural language processing, and bots, to name a few.

Yet there are reliable studies going back in marketing on the use of neural networks in general on segmentation, targeting, and positioning, and these studies are of more interest to the business and marketing professionals. Early applications of neural networks focused on the use of unsupervised learning where the input and output are segmentation criterion with a single hidden layer, and the goal is to predict membership to a segment. These methods were reliably found and reported to outperform the traditional k-means clustering approach since 1999.

Similar early applications can be found in the area of order fulfillment and inventory management where piecemeal applications of neural networks have enabled inventory optimization, managed reorder levels, and been deployed for route optimization, among others.
These methods and concepts are now giving way to more integrated solutions where the networks are becoming more sophisticated and connect to one another, multiplying the value by feeding information across the entire value chain.

Neural networks have also consistently outperformed traditional multinomial models when it comes to lead qualification and targeting while enabling companies to seamlessly scale these results in real-time to millions of customers. For the past decade, researchers have reported findings from studies on neural networks outperforming logit models in direct mail and marketing campaigns. Neural networks, in combination with evolutionary selection algorithms, have been found to be much better than traditional principal components types of methods in identifying the subset of features that maximize classification accuracy. These approaches are increasingly being used to do real-time targeting, leveraging a combination of website activity, streaming data, and social media in applications as diverse as security and safety to targeting of offers through mobile applications increasingly impacting the world of big unstructured data problems. Chatbots are now further providing the capability to engage customers in conversations through the adoption of deep learning techniques in place of the earlier Markov chain-based approaches. These approaches are in the early stages of adoption with immersive technologies like mobile virtual reality (VR) and augmented virtual reality (AVR), now beginning to enter the market to allow for more immersive presentations and virtual experiences. For fans of the 1987 Robocop movie, there is an early analog in the K5 robot from Knightscope, which is billed as an autonomous data machine and can patrol malls and campuses providing much needed physical deterrence on the cheap.

As applications of these methods expand, and computer power becomes more available, we can expect to find increasing applications of these methods in the physical world. The IoT space promises good dividend by combining the newest concepts in computer vision and deep networks in allowing seamless automation in diverse areas from autonomous vehicles, robotic manufacturing and assembly, retail shopping experiences, and automated kitchens, to name a few. As we look into the future, we can expect single-point solutions to get replaced with increasingly complex combinations of capabilities across methods and fields to be put together to deliver more robust solutions.