Augmented Stylist

Idée fixe came to my head during a Black Friday online shopping session while I assembled the winter season’s wardrobe, so it’s no accident that we used fashion as an example. As you will see, this example is almost perfect… but let’s not be so hasty.

Mysterious stylist of mine

After picking a few clothes that were my primary need, I started to research the website looking for ideas (and obviously for more, even hotter deals), and here where something weird happened.

Suddenly, I noticed that I started to research light jackets, despite the same day ground frost hit for the first time. Jackets were not included in the Black Friday sales offer either. I found out quickly that my shopping path had been changed by nothing else than quite a nice outfit proposition composed by brand stylists, nicely photographed, and included as item imagery. Wow.

Simultaneously, I was writing a comprehensive guide for recommendation strategies, with an emphasis on AI-based recommendations. Two questions crystallized in my head as a result:

Does human, experience-based recommendation still outmatch artificial ones?

Who will be the better salesperson while composing a wardrobe—an AI-recommender or brand stylist, designing outfits for photo sessions and overall brand-style?

If AI, which strategy of recommendation is most suitable for fashion? If it’s the stylist, why is quite an outdated marketing approach—One-to-Many— still better in practice? Style is all about being unique, so why is one proposition winning this battle? And here comes the second question:

Who actually drives the trends?

Is there a chance that AI-boosted recommendations are creating their own trends? Or bias it somehow, or at least, influence it? At the end of the day, trendsetting is all about biasing and influencing present styles. Is fashion, as one of the aspects of art—humanity’s last stand—slowly falling into Artificial Intelligence grasp?

The answer isn’t simple or straightforward and requires an approach from many angles.

Recommendation as a conversation

The first, most basic, division of recommendations is between explicit and implicit. Explicit are choices we make consciously. “Are you looking for men’s or women’s pants?” Based on our declaration, the algorithm selects an offer.

Much more useful are implicit recommendations, in which we don’t straightforwardly declare our preferences but rather reveal them. They are more effective because, quoting folk wisdom, “actions speak louder than words.” In this case, we can take this proverb quite literally.

AI recommendations, which we will discuss, belong mostly to the latter category.

Recommendation as a filter

To be precise, in general, recommendations are an aid for proper customer care that cannot be overestimated. Even straightforward ones like the Recently Viewed Products can help the shopping sessions to finalize.

This approach is called segment-of-one or, more formally, Machine-to-one. The brand is represented by automation to communicate with clients with such dedication, which creates a feeling like every brand’s content is tailored.

It replaced outdated approaches like One-to-many, which I mentioned above. Every client is different, and by default, has different expectations, needs, and – especially – tastes. Thus, presenting the same set of items is risky and lower converting than dynamic client-oriented offers. Pretty obvious, isn’t it? What’s more, such a recommendation comes with various types and strategies.

• Recommendations based on various inputs can recommend different items. View-based, e.g., will be entirely other than purchase-based.
• Fashion speaking; In various levels of it, other factors matter the most, yet Gucci, Off-white, and Primark are advertised differently, aren’t they? The same rule applies regarding recommendation strategies.

One of the most popular recommendation engines is Collaborative Filtering and isn’t reserved only for the fashion industry. It processes the most important events the client can perform – product views, adding to cart, and purchase.

The algorithm records these actions and proposes products based on a behavioral model. If one model is similar to another, it suggests products. Do you recall the “Look what we’ve prepared for you” frames you can find while purchasing stuff online? In most cases, this is a collaborative filtering-based frame.

What about the “You may also like” frame at checkout? It is a frame reinforced with Market Basket Analysis – another recommendation golden standard. While Affinity Analysis is based on the general behavioral model, Market Basket Analysis (in broader perspective called Affinity Analysis) focuses on clients’ actual choices – exclusively purchases. Affinity is measured between products often bought together.

It’s not that important how exactly Affinity is counted right now. It’s important that the relationship between two items might be expressed as a number. Speaking of Market Basket Analysis used in the edrone system; the lower it is (less than 1), the bigger chance is that one product replaces the other in practice. The higher (greater than one), the greater the probability that the products somehow extend each other.

If you find it similar to the two quite often used techniques: cross-selling and upselling, you’re totally correct. You can treat Affinity Analisis as cross-sell/up-sell on steroids. Classical techniques augmented by Artificial Intelligence.

As you see, commonly used techniques are based on customer choices. Does this mean that AI-based recommendations, at the end of the day, are actually reflecting human-generated trends?

Recommendations as trend-conducting medium

The activities mentioned above work the best with almost all types of clients (especially clumsy ones, jumping from site to site, and easy to be distracted.) But it’s hard to call it a trend generator. Better to say a trend exchange medium.

The role of AI for recommending the items is to successfully predict client choices based on their unique tastes.

1. Repetitive
2. Driven by availability
3. Driven by fashion trends

In theory: also driven by their preferences, but in practice, it is also dependent on the three factors above. Things are getting complicated… Let’s put it into lemon tones. It’s the graph I prepared while researching for this article.

Not quite explanatory? Well, it just barely covers the main dependencies between factors. But if you give it a second try, it starts to make sense. The network of relations, inspirations, propositions, and possibilities is vast. Complex and filled with feedback.

Things are getting weirder if we will take into account yet another factor mentioned above.

Recommenders already bias the style choices! AI can also generate trends themselves.

AI choices are backed with input data, but output has an element of randomness in it. Each AI verdict is a list of probabilities of the correctness of each of them. If, as a result, we have a few suggestions, there will also be those less accurate… which, unexpectedly, may appeal to the client, increasing their chances of appearing in the recommendations again.

Moreover, this error margin is an essential part of recommender system operation (and other machine learning algorithms). If we force the machine to always give 100% accurate results, we need to train it to death, using a small amount of training data (it’s the only way to achieve that in a reasonable amount of time. In overfitting (“overlearning”), the system loses its ability to generalize and adapt to reality.

If it encounters input that differs from that on which the system was trained, it cannot give any output. Overlearning causes it to lose its intelligence.

Who is the true elegant arbiter in a world dipped in postmodernism?

There is no one right answer, as I mentioned in the introduction to this article. As I went deeper and deeper into my deliberations, it became clear that the more the world becomes a global village, something like style becomes an extremely fuzzy concept.

It is influenced by practically everything. We are surrounded by access to any fashion level, whether we are talking about products or inspiration. The inspirations themselves come from newer and newer places, we also have more freedom in matching elements of this style with each other – we allow ourselves to make more and more unobvious combinations, more and more often coming from different periods and different trends in history.

The role of artificial intelligence cannot be ignored. Recommendations based on AI are catalysts, transmitters of choices, allowing for the remote transmission of style without entering into a relationship with an unconsciously inspired person.

One last thought

We mentioned explicit and implicit recommendations. I raise the bet. What if – considering fashion, we are under the effect of an even stronger and frightening omnipotent force – social media. Eventually, we are all trapped within our own information bubble, thus truly a fresh inspiration, let’s say, out of the box, has the much harder task to reach us. Usually, we are inspired by people we admire, or at least respect, and something that is abstract, from the point of view our fashion discussion, like the newsfeed’s recommendations system, may have a much bigger impact than the most sophisticated recommendation engine.