I am currently taking part in the KPMG ideation challenge where we were tasked to use artificial intelligence to change the way a business services their customer base.
The initial pitch
We want to introduce a predictive tool for targeted marketing, it would initially be optimized for the retail industry but as the insights from the tool improved it could be applied to other industries. Targeted advertising today is largely based on search queries among other metrics. It works well and is a multibillion dollar industry, in the first quarter of 2015 google took in 15.5 billion dollars. It relies on people to input information of things they are looking for. What we are suggesting is to take this a step further, what if we can analyze what a user is likely to buy before they look for it. As an example by analyzing a user’s Instagram posts we can see what the clothes they are most proud of and what they think they look best in. Additionally, unlike search queries, this is clothes they have already spent money on, not clothes they could just be looking at. By using this information you would gain the competitive advantage of better insight into what the user is likely to use and you get it before the competitors relying on the user to search for it first. It would take targeted advertising to predictive targeted advertising. A system that knows what you want before you do and eventually better than you do.
How would we execute?
The development of such a tool would come in stages.
- We would train an AI to use image recognition to distinguish between different articles of clothing and accessories that people wear. We would simply piggy back one of the Keras AI’s that already are primed for image recognition and reconfigure it for our specific images that we will train using images scraped from the web. The longest part would be the scraping of images but since the AI is already primed we will require less images than an AI from scratch would. About 1000 tagged images for training and testing should be enough. This would take about 40 hours to scrape and 30 hours to code.
- Once this is perfected we would group different clothing articles together to form an idea of the type of clothes the person in a picture wears. I.e colour pallet, style, type of clothing. It could then AB test ads and get better at figuring out what ads to send to which people. This step could take any number of hours depending on how many times we iterate, an educated guess would be around 30 hours using a script and data analysis. In theory this could also be done with an AI and as more people are analyzed and more advertising is done, it would create a positive feedback loop that would perpetually produce better results. This technology could either be sold directly to retailers as a tool for them to gather insights. Or, users could be targeted directly off our platform in the way that google adsense or facebook pixel do. The advantage of the former technique would be that this is more enticing for a retailer and it would save us from the brunt of privacy concerns. The advantage of the latter is that if we control the data, the AI would get better and it would be the most lucrative in the long term. It would just be harder to launch in terms of resources and would take longer for a return.
Updates to the plan
Fashion is ever-changing and incredibly subjective. It would require a lot of work to try and come up with categories for clothing styles and would be incredibly difficult to do accurately. It is because of this that we decided to change tactics and use unsupervised learning methods for the categorization of clothing. This would take a lot longer to code properly since we wouldn’t be giving it any discrete features. However we decided that this route was worth it because we get the benefit of scalability. Refraining from explicitly classifying and labeling new styles would allow for a dynamic model that would also be more accurate. It is the only viable way of making a product on scale.
Problems moving forward
There are a few problems with our current pitch that we still need to figure out.
- Privacy concerns/ data – we can’t mine people’s data from their social media as it probably hits some privacy laws.
- A company that has user data can just build this themselves – we can’t use data from a company like Facebook or twitter because that’s the data that makes their company valuable, it makes no sense for them to allow us to build such a product using their data.
- How do we make money – we have to finalize the business; do we make a tool to sell to retailers to better target their own customers? Do we sell the ads and target the users ourselves?
Some ideas for these problems could involve setting up cameras in store and detecting what you can off of what people wear to your store, building up a database of different customers by using facial recognition to assign ids to them. Or maybe we could leverage mailing lists retailers have and try to reverse search them to build a profile of their customers. Whatever the solution is, our biggest concerns always circle back to privacy concerns. If we are to move on in this competition, it is our biggest hurdle.