While the technology behind GANs is complex, the concept is quite simple. A GAN is comprised of two neural networks (a neural network is a form of AI that is good at recognizing patterns). One neural network is called the generator. The other neural network is called the discriminator. Both neural networks are trained with the same data set, but have different tasks. The generator creates new content based on the similarities it finds in the data set. The discriminator, on the other hand, compares the new content against the data set to see if it is good enough to pass off as real. This constant battle between the neural networks is what makes generative adversarial networks ‘adversarial.’
Not only are GANs dynamic, but they are also adaptable. Once trained, they do not stay static. The generator continually refines its creations to try to increase the portion that can slip past the discriminator. The discriminator, conversely, improves on its methods to catch even more of the generator’s creations. GANs operate like a continual cat-and-mouse game between a counterfeiter and a cop. Furthermore, new, raw data could be added to the original data set to make the GANs generate and discriminate new information. This adaptability means that GANs never have to become out-of-date.
Specifically, GANs will play a critical role in marketing. Even the greatest products and services can fall flat with poor marketing. Businesses can employ GANs as a form of test marketing to predict and improve how consumers respond to new products, helping them overcome the challenge of marketing and put their products/services under the best light possible.
There are multiple ways that GANs can help with marketing. For one, businesses could employ them to improve product recommendations in online marketplaces. In January of 2018, three researchers at Amazon India Machine Learning trained the ecommerceGAN (ecGAN) to do just this. Theoretically, a customer could buy any combination of products from an online marketplace in the same purchase order. The ecGAN explores the space of all possible orders a customer could make and determines which orders a customer would actually make. The generator creates possible order combinations while the discriminator decides if that order is realistic. The original training data set of the GANs can also be updated to include new, arising trends of the digital marketplace, allowing GANs to recommend products that both today’s and tomorrow’s customers will actually want to buy together. This new model tightens the accuracy of product recommendations on websites.
In addition to being involved in advertising, GANs can also help with finding information about sales. The same three researchers at Amazon India Machine Learning developed the ecommerce-conditional-GAN (ec2GAN) to predict the demographics of the customer, selling price, and the general dates for sales. As it turns out, the ec2GAN could make accurate predictions with its generated scenarios that lined up with the real-world data. For example, products characterized as men’s shorts were predicted to be sold 63.10% percent of time during summer months. Real-world men’s shorts were purchased 67.52 percent of the time during summer months. When used across customer demographics, prices, and seasons, the ec2GAN can provide invaluable information for retailers.
Knowing what is bought together, when, for how much, and by who is a good starting place for marketing information; still, it is important to know how the product listing could be changed to influence sales. In an increasingly digital marketplace, product descriptions play a crucial role in selling the product. Two researchers from Stanford trained a GAN using information from Airbnb listings in Manhattan, New York between January 1, 2016 and January 1st, 2017. In total, 40,000 listings were used to train their GAN. Supposedly, if a listing had a better description to increase its appeal, then it would have a higher occupancy rate. The hypothesis proved false. The description did not have as large of an impact on occupancy rate as other key variables like location, home type, and amenities. The exception, however, was keyword packing. Listings that used a lot of key search terms became more visible on the website, and thus had higher occupancy rates. GAN technology was able to identify this significant pattern for businesses to take advantage of.
GAN technology can be applied to gather useful marketing information to aid businesses. Despite having been just coded in 2014, the technology has made great improvements in the past five years. As researchers continue to refine different types of GANs and look for more applications, it is clear that GANs are here to stay and they will only find a larger place in the business world in the years to come.