This post was originally published on Forbes.com.
Personalisation has become one of the “must-haves” for online shops in recent years. According to research by Epsilon and GBH, 80% of US online shoppers are more likely to make a purchase if retailers offer personalisation. This could be in the form of a product recommendation, shown as “you might also like,” or special email offers with curated content.
A big advantage of personalisation is that it helps shoppers navigate overwhelming e-commerce offerings and get to the purchase decision faster. Who wouldn’t prefer to go to an online shop and immediately find exactly what they need, instead of spending hours on browsing or doing research? But it is not just about making quicker decisions. It can be critical for a business to help shoppers overcome “analysis paralysis”, a phenomenon where too many options leave one unable to make any decision. Accenture found a worrying example of this in their research – “nearly 40% of consumers have left a website because they were overwhelmed by too many options”.
Fashion, with its many sizes, collections, and styles, is a lot more complicated and overwhelming to consumers than most other retail categories. Despite an iPhone being infinitely more complex product than a piece of clothing, choosing which model to get is still an easier task than figuring out what shirt will work best with that new pair of pants that you’re buying.
With all these complexities of the industry it is no surprise that conventional approaches to personalisation are falling short in fashion.
Challenge 1: Collecting the data
One of the fundamentals of any personalisation effort is gathering data. Not just any data, but data that is reflective of target consumers’ habits and can be representative of their future behaviour. It’s also important that there is enough of it to account for statistical deviations and variations.
Creating a data pipeline that enables a company to gather this data regularly and effortlessly is a challenge on its own. One way to approach this is to make the data pipeline a part of your core product and shopping experience. The US and UK based online personal styling service Stitch Fix is doing just that. They create curated fashion experiences where shoppers answer questions about their preferences and expectations before getting matched with a stylist who creates a “fix” (a set of five items) that they think would work for the shopper.
To fuel the technology, they rely on customers sharing useful information from the very first interaction. Shoppers fill out a quiz and tell them about fit, size, style, and budget preferences and any additional information that would be useful for a stylist. They also share more personal details that can be hard to capture in a regular online shopping experience – for example, whether they have a holiday or special occasion coming up, or are looking for a particular floral dress.
Stitch Fix’s business model is built in a way that this actionable and granular data is not only gathered once but is continuously updated. Every time a shopper keeps or returns an item, more valuable insights are added – for example, that a particular item was too big or a shopper didn’t like the style. These insights are carefully recorded and used to better understand what a particular shopper will like next time. The company also uses Style Shuffle, a Tinder-like product where customers can swipe right to styles they like. Thanks to this sophisticated, yet effortless (for consumers) data pipeline, Stitch Fix is able to generate a unique “style map” that visually represents the range of styles shoppers are most and least likely to love. Most importantly, this data is updated, every time a shopper interacts with the services, ensuring that the company’s understanding of their style and preferences is up to date.
Challenge 2: Is the data still relevant?
Personalisation technologies are usually fuelled by massive amounts of data about shoppers. The idea is that one needs to have a substantial amount of consumers and data about their behaviour before start predicting what they will like.
At least that is what used to be the norm in the industry. I spoke to Anna Kuragina about this mindset. She is the Product Area Manager for consumer facing artificial intelligence products at H&M Group and she said: “Conventional approaches to personalisation and product recommendations utilise a long history of customer data to predict their future needs and behaviours. Typically, in those approaches the rule is ‘The more we know about our customers from the past, the better predictions we get.’”
But simply gathering every piece of data at all times and analysing it uncritically is a naive approach and can lead to its own problems, something that the Swedish fashion giant is well aware of. Anna went on to add: “We live in a rapidly changing and unpredictable world where customer preferences and fashion trends change quickly. We can’t afford being biased by what happened one year ago, often even one month ago.”
This is especially true, when brands operate in a fast fashion cycle, manufacturing new collections many times per year. Data about consumers is also much more likely to decay in the fashion industry than most other industries. To put it simply – one’s body size or style preferences might have changed significantly since the data has been collected. In this high paced environment, staying on your toes and capturing the most recent insights is paramount.
Challenge 3: When no two items are the same
Utilising data about products that are the same or somewhat similar is another common approach in personalisation technology. If the same shirt has been bought by hundreds of shoppers, it’s going to be easier to predict how others will buy it. But what do you do when every item you sell is unique? This is exactly one of the challenges that Vinokilo, an online platform for vintage fashion, is dealing with. On their website, there is no variety of possible sizes or colours for an item – only the one individual item listed.
Typically fashion is more standardised and substitution is easier when there are hundreds of products in dozens of collections from thousands of brands. But vintage clothing has a lot of variety without the abundance. It becomes harder to nail down why a shopper liked an item – was it because of the brand, its retro print, or maybe because of a distinct 90’s feel? When you have a sample size of one it becomes extremely hard to make any kind of prediction. It also leaves less data for the algorithm to learn from, because choices like multiple sizes or alternative colors are not there.
This makes creating a satisfying shopping experience more challenging than it is for regular fashion products. If a shopper finds an item that they like, but it is in the wrong size or already sold out, that shopper is left disappointed and with no options other than to move on because there was only ever one of those items for sale.
I talked to Anisah Osman Britton, the Chief Technology Officer at Vinokilo and she told me about their approach. She explained that they try to learn from each unique piece of vintage clothing and understand what other items might work for a shopper. For example, by looking at the vintage product as a combination of factors: brand, style, category, quality and others.
The company is currently piloting a feature where their top customers are notified about items that should match their style, before the items even show up online. This is useful because there is always a delay between the time they receive an item in their warehouse and when it appears online. Additionally, because there is only one of each item, matching such items with the right shoppers makes it more likely that they end up where it will be appreciated the most.
This is no easy achievement and doesn’t happen by itself. The company combines algorithms with manual data entry at their warehouses to ensure that they can gather the needed level of granular data about all the items that come through the supply chain.
Challenge 4: When algorithms alone can’t do the job
For some industries the challenges of personalisation simply come down to collecting more data and building more accurate algorithms. But in the world of fashion, the algorithms and data alone might not be enough. Figuring out someone’s individual style and perception of what looks good can be a hard nut to crack just by relying on machine intelligence.
One way to approach the challenge is to let algorithms do what they’re good at and leave questions of taste and style to humans.
For example, Stitch Fix matches shoppers with stylists to ensure that items selected for a shopper are the perfect match. Not only does the company create a “style map” representing the shopper's preferences, they also do a similar analysis of the stylist's own style. “We are able to bring these two sets of information together and use them as the basis to match the client with the best Stylist for them from the outset”, they explained and added, “the stylist’s relationship with the client is not only key to the overall client experience, but also adds a critical layer to our data feedback loop, helping to fuel Stitch Fix’s personalisation ecosystem.”
With stylists playing such a central role in building strong relationships with their clients, it’s important that they can focus on things that humans are better at like communicating and building relationships. Or to interpret nuances in clients’ unique requests like “I need a killer dress for a wedding, and it needs to be below the knee”. The team at Stitch Fix also told me that the company has started using Natural Language Processing of the clients’ requests to filter the initial products that the stylists work with. This leaves the stylists with more time that they can spend on making stylistic decisions rather than manually remembering to remove pink shirts from the fix because the client said they don’t like the colour.
Another example of combining the human touch with algorithms is Anomalie, which sells custom wedding dresses online. It is hard to imagine a piece of clothing with more pressure on it to be absolutely perfect and personal than the wedding dress. It would be similarly hard to imagine that such an important dress could be successfully bought online rather than in a physical wedding dress store.
Anomalie provides brides-to-be with a tool for creating customised wedding dresses from different starting points - like choosing a neckline, silhouette, or length or by browsing sketches.
To learn more about their approach I interviewed Gillian Langor, the Director of Fashion Technology at Anomalie. She explained that because wedding dresses have such a high bar and need to be perfect for a bride, the company has two key goals. Firstly, to fully understand what the bride wants. Secondly, to match the bride’s expectations with the shopping experience. To achieve this the company combines collaborative filtering, a method widely used in personalisation technology, with a team of in-house stylists.
Popularised by companies like Amazon and Netflix, collaborative filtering looks into consumers with similar habits and tries to predict what other consumers are likely to do. At Anomalie they try to find other brides similar to the shopping bride in order to give better suggestions based on shoppers’ preferences. For example, what style would work best for the bride or find the perfect “nude” shade of the illusion mesh, that is often used for lace applique or beading. A “nude” colour means many things depending on your skin tone, and often darker skin tones are not supported in the same way. The company addresses that by looking at what brides with a similar skin tone bought and what would look good on the new bride.
For Anomalie combining algorithms with the human touch means finding the important balance between guiding a bride and providing all available options, while using their best professional judgement.
Challenge 5: When the Netflix approach falls short
While collaborative filtering is often fundamental in personalisation technologies, this approach can sometimes fall short for fashion. Anna Kuragina of H&M Group says, “Compared to other retail players, e.g. grocery retailers with more permanent assortment, the fashion industry is facing an interesting challenge. How can we use customer data to be relevant while still surprising our customers with items that would speak to them right at this moment?” She tells me that at H&M Group, they decided to move away from traditional algorithms such as collaborative filtering and “become creative enough to build custom-made models that will help predict the fashion our customers desire before they even know it themselves.”
The shortcomings of the “Netflix” approach to recommendations also become obvious when dealing with the challenge of size and fit. With sizes differing so much between brands and even within one brand, there are no easy solutions. At the company I founded, Easysize, we have created an AI solution that recommends the right size and fit to online shoppers without using body measurements or size charts. I recently talked to the CTO of the company, David Babayan, about some of the challenges of dealing with the variations of fashion products.
“It’s important to understand the complexity and the variety of factors that play a role in finding the right size. From the clothing cut and fabrics, to the customer’s preferred style. It is naive to approach the topic as a simple issue with measurements or a categorisation problem. Because products that have the same size or measurements, can have distinctly different styles and therefore fit and feel,” he explained.
For example, instead of simply treating the fabric of the item as a categorical feature, a factor representing “stretchiness/elasticity of a fabric” is used. This property accounts for the changes an item undergoes over time and leads to much better predictions. This is something that experienced stylists and shoppers always take into account, so it’s been incorporated into the size recommendation algorithm as well.
Another important factor is understanding how customers will be wearing an item. The same shirt in the same size can be worn by three different people with different bodies depending on their preferred fit and style. But the same shirt can also be worn by the same customer in three different sizes and ways as David says, “A person can provide you with their exact measurements but still prefer to wear some types of clothing looser or tighter than the basic size chart accounts for.” Maybe the shirt is worn on its own for a tighter fit for work, more casually under a sweater or perhaps on top of a t-shirt, like an overshirt. It’s still the same shirt, but there’s three different uses for it.
Where do we go from here?
There is no doubt that personalisation technologies already play and will continue to play a big role in fashion e-commerce, an industry that is, by its nature, very complex. People don’t just buy fashion for utilitarian reasons, but to feel good and look a certain way. This behaviour is often driven by a lot of factors that might not be obvious from the first glance. These variables can be hard for humans to pick up and maybe even harder for an algorithm to discover.
For example, how can an algorithm know what exactly it is customers like about a particular item? Was it the colour, or the fabric, or because they saw a similar dress on Instagram? Can an algorithm detect what accounts for the changes in behaviour when shoppers buy from a flash-sale site where everything is sold at 70% discount compared to when they buy directly from a brand? How can an algorithm address the constant cold start problem, when there is simply not enough initial data, when fashion constantly has new brands, new items, unique items?
As an industry, we are still far from having the ideal personalisation technology that will not only account for all the requirements a shopper might have but also fully understand what shopper’s perceived style and expectations are. Conventional approaches to personalisation must be modified and adjusted to account for the industry’s unique challenges.
It is a hard task, one that is likely to be solved by applying a generous level of creativity and taking into account all of the small irrational habits of customers and not just like a straightforward math problem with logical conclusions. After all buying a light floral dress for summer holidays might not be exactly the same as buying a new iPhone.