It can be really confusing to know which are the most beneficial products to buy if you are into make-up and skincare.
Is there really any difference between a BB cream and a CC cream? How essential is the toner part of a cleansing routine if you have dry skin? What is the best moisturiser to use if you have oily skin and are travelling to a humid climate?
If consumers are asking themselves these questions, brands and retailers are also pondering the same thing as they push to personalise the user experience that they offer.
With the half decade’s worth of beauty data generated from a platform that generates shopper profiles and recommends beauty products, CEO Nidhima Kohli has recently uncovered Beauty Matching Engine. According to Kohli, Beauty Matching Engine helps brands, hair salons and beauty retailers to personalise the shopping experiences of their customers by tailoring their recommendations, upsell suggestions, emails and landing pages.
All of this is powered by data but, as previously mentioned, products in the beauty industry can be perplexing so gathering all of that data to power the engine initially proved to be a headache for Kohli.
“The hardest bit about using artificial intelligence in the beauty space is the taxonomy. I could have five different red nail polishes and none of them would be ’Chanel red’. They have names like ‘glaze’ or ‘midnight star’ so you cannot have a colour taxonomy. It’s a mess,” she said.
Another teething challenge was knowing where to start to collect the data with so many products available that seem to serve the same purpose. “When we were launching with one of our first clients, just looking at moisturisers and only two or three skin concerns, we had more than 900 user personas,” said Kohli.
"You need the confidence level to be strong enough to make those assumptions."
She also said that getting the critical mass of data was important. “You need to have the statistical significance so the sample size is big enough so the confidence interval is strong enough and relevant enough to be able to make those assumptions and machine learn.”
"We cannot get this wrong."
Furthermore, the stakes are far higher when personalising product recommendations for beauty and skincare products than clothes for example because of the biology involved. Kohli said: “We cannot get this wrong.”
In terms of how the recommendations are figured out, Kohli explained that she and her team of data scientists, who use R and Python, make use of reinforcement learning, in the manner of Netflix. For example, someone concerned about acne is shopping for an anti-wrinkle face cream on retailers’ website.
“We measure propensity to sell and that gets re-fed into our engine."
They are offered five products. If three other customers with the same concerns also browse for a face cream, Beauty Matching Engine will look at the markers in their profiles (such as age, skin type and climate) of the other people browsing and see which one actually does click the ‘buy’ button.
“We measure propensity to sell and that gets re-fed into our engine and we might learn that out of these five products, skin tone doesn’t affect propensity to purchase as much as age, because the younger you are, the more you are willing to spend on an acne-treatment product,” Kohli explained.
She found that it didn’t matter too much that her data scientists were not domain experts with experience in the beauty industry, although she did have to explain and document how ‘weightings’ work.
This is where factor A will have a bigger effect on a final result than factor B. Kohli illustrated this with the example of eczema. “If a person has eczema, it might affect what shampoo they buy but not what nail clippers they buy,” she said.
The B2C white-label platform is making an impact with nearly 18 million recommendations and conversion rates of between 30% and 300%, depending on the retailer and the brand. Going forward she is looking to implement the add-on of a data and analytics dashboard to the backend to enable clients to view browsing rates versus buying rates.
So why was Kohli best placed to fix this problem? Two reasons. She has the accumulated data to help her do so – from historical from the first platform, and current Google Analytics, retailer or brand data, competitor intelligence data, as well as weather, pollution, UV levels, IP address and customer location data.
Secondly, it could be done. “I was seeing that personalisation was working very well in fashion. Lots of people have been talking about it for beauty and no one has been doing it,” she said.