Escaping the ‘Filter Bubble’ of inaccurate recommendations:
Navigating product recommendation systems in fashion e-commerce:

By dr. Erik van Breusegem, Head of AI at PTTRNS.ai

Product recommendation systems - a valuable ally

Most e-commerce brands and retailers today offer some kind of recommendation engine and functionality on their platforms and channels, for good reason. It is hard to overemphasise the economic value of recommendation tools, even though quantifying their exact impact can still be difficult. Indeed, a well-executed recommender system can drive uplifts in conversions by multiples, improve basket size, and reduce returns, all reflecting a happy helped customer. Even simple ones that use off-the-shelf AI tools can yield surprisingly positive results. No wonder most apparel brands and retailers usually have multiple recommendation mechanisms active. Many common uses include offering a jump-off point as assistance with navigating at the start of the customer interaction or acting as reassurance that the product they are considering is the right choice, later down the line.

(Not) bridging the trust gap 

Getting the maximum value out of a recommendation system, however, in practice is easier said than done. In particular, a typical online environment struggles with a lack of relevant consumer data to start off and a very limited time to provide valuable assistance and build trust. This is sometimes referred to as the cold start problem.

The cold start problem is common, especially in fashion retail since generally accessible parameters at first entry (like software installed, used, cookie information, etc.) disclose little about a person's relevant characteristics. In the future, the challenge of staying immediately relevant will only intensify with stronger privacy protections and stringent regulations like GDPR (and related), Cookie laws, and new AI-related laws in the EU.

 Most brands solve this problem by not solving it. For instance, they'll wait until your customer starts navigating by themselves before they start recommending or will start with a relatively safe, (pseudo-random) recommendation to see how the visitor reacts.

 While measurable value can be captured with such a strategy, it won’t build loyal customer relationships, and leaves the customer in a transactional mode. Both the retailer and the customer end up in a situation where the focus is on getting to the quick purchase, rather than building lasting relationships, trust and mutual understanding. This trust gap leaves much potential value on the table for both the retailer and the customer, leading to higher product returns, poorer loyalty, lower product satisfaction and higher acquisition costs in the long term compared to a more relational approach.

The feedback loop in recommendation systems

The standard transactional way to operate in addition tends to also create irrelevant filter bubbles, further widening this trust gap. Why this is happening requires a bit of “back to basics” – how current transactional recommendation systems tend to work.

In recommendation systems, when a customer is presented with certain options, the decision they take on whether to interact with a provided aspect is passed (fed) back into the recommendation model. This establishes a feedback loop that impacts the user's future recommendations, often by showing more of what they have interacted with previously. This reinforces a consumer profile that is based on positive interactions (e.g., likes, clicks, and adding items to shopping baskets, etc.). In comparison, negative or absent interactions have much less of an impact on this perspective. Imagine browsing a fashion e-commerce site and clicking on (interacting with) a black bag. As a result, the shopper will receive suggestions for other similar items, such as more black bags. In addition, to scale up the recommendations, the feedback loop must also be automated. While using an automated feedback loop is a valid approach and has its benefits, it does come with its own set of challenges. Particularly relevant for the transactional vs. relational model is the risk of getting trapped in an irrelevant "Filter Bubble."

The ‘Filter Bubble’ effect

In the context of fashion brands interacting with e-commerce visitors, there is a practical challenge that often occurs: the customer ends up in a narrow dead end of only seeing the same product with the same attributes. To use the same example as above, If a visitor is interested for a brief second and clicks on a black bag, they’ll likely be led and get stuck in a stream of exploring only the same or very similar products, although their needs might have shifted or they were looking for a completely different product altogether (e.g. because of a misclick).

Filter bubbles tend to happen quite naturally in an unmonitored fully automated implementation of a recommendation model. One important dynamic to highlight is that a positive interest signal, such as clicking on something, holds much more value in such systems than a negative one like not clicking. This is because the possibilities of what you can't do on an e-commerce site are an order of magnitude greater than the interactions you can actually make. Consequently, anytime a consumer interacts with a product, filter, or anything else on a website, it tends to drive a huge boost to a recommendation model that is automated.

Another driver is the limited screen space, particularly on mobile devices, and the need to reach decisions fast to avoid losing the customer's interest (in the traditional transactional interaction model). In the worst case, this can lead a recommender to quickly focus on a few related items that are unsuitable for the customer in question, with no way for the customer to indicate that they have been led to the wrong place, or even that their intentions have shifted. In this case, the only way out then is to close the session and start completely from scratch - a very unsatisfying outcome, as it may just as easily cause a visitor to give up or go elsewhere.

How to prevent an irrelevant ‘Filter Bubble’ – quick interventions

1)  Offer multiple recommendation options, not just single matches      

The primary principle is to constantly provide customers with several recommendations visible, ensuring minimum amount of diversity. A good rule of thumb is about five. This provides at least some freedom to navigate away from certain items and helps to avoid getting stuck on the same product or product attributes. It's also a natural extension of the fashion shopping process, where the focus is on exploring and discovering a selection of great choices, rather than determining an objective ‘best’ option.

2)  Have one or two odd recommendation options

Another alternative, which may seem paradoxical, is to specifically present one or two products (out of the five suggestions) whose qualities substantially differ from what is currently thought to be the best choice range for the buyer. This offers a safety valve since we could have misunderstood the customer after all, or a mind-shift could have occurred (e.g., shopping now for a different occasion or a friend). What’s important in this optimisation is to be explicit about why the different option is presented. One natural method to introduce this concept is to present a question to the customer such as for example: ‘Are you looking for something different after all?’ or ‘Feeling adventurous, maybe consider this instead?’
The intent is dual: to get a signal that we might be heading for an irrelevant bubble, while at the same time gaining more information on what specific style element or product attribute the customer is looking for.

In addition, sometimes acknowledging failure can be just as beneficial. Although unfortunate, this is an unavoidable outcome when relying on information such as prior history or factual information (e.g. physical address), since that information could have lost its relevance or become plain wrong. Therefore, a good practice is to run a “dissonance” detector in the background. In this case, if the actual behaviour of a visitor is not congruent with the recommendation, or if the information itself is contradictory, reassessment is warranted. This requires, however, rather strong data analysis and AI techniques in the background.

Avoiding irrelevant recommendations with a relational approach

While the above adjustments can help, as long as we stay in the transactional approach the irrelevant filter bubble risk will remain quite high. Eliminating this risk in our experience mandates a deeper relational approach to recommending.

The chances of getting into an irrelevant filter bubble if you take a more relational approach are far less likely, since you have a richer and fuller view of your customer and only go to definite conclusions once they are justified. This means starting with gathering deeply relevant information from online customers first.

Well-known long-term subscription-based platforms such as Netflix and Spotify, for example, have managed to build strong customer relationships by gathering hyper-personal data from each user. This is easier to achieve compared to fashion retail experience since customers interact with those platforms daily. When it comes to fashion e-commerce platforms, however, customer engagements occur much less frequently, since shopping for fashion and apparel comparatively is a more occasional activity, whereas listening to music often is a daily and continuous interaction. In addition, shopping for clothing is often influenced by a range of personal, situational, and environmental factors, making it even more context-driven than for example, listening to music. Factors that influence a customer’s need or want to purchase from a fashion brand can be personal expression preferences, occasion, environment, fit of the clothing, seasonal changes, lifestyle and budget, sustainability ethics, etc. In this sense, shopping for clothing is more diverse and situational (context-driven), which makes implementing personalised recommendations at scale and organising a relational strategy a bit more difficult.

Few fashion retailers and brands therefore aim for a relational approach, either because they fear that visitors will lose interest or simply because it seems too difficult to create. Today’s data-driven (AI) technologies, however, now allow for a more personalised relational approach in a scalable way, and much more accessible.

Enabling a relational approach in fashion e-commerce

A solid recommendations system built on relational principles should address the customer’s underlying psychological needs to achieve longer-term brand trust and engagement.

In fashion and apparel commerce, for instance, relevant psychological demands to discover could be: Does the customer want to feel more confident or empowered? Are they looking for a unique piece - a sense of individuality or perhaps they are following the trends - a sense of belonging, etc?

A way to address those needs can include helping your clients overcome the paradox of choice and making them feel understood and listened to. What this means for your customers is having a genuine advisor by their side who offers exclusive help on what is the best product for their needs. While relatively new in an online setting, personal shopping assistance has been used for a long time in a physical retail environment. In the real world, fashion stylists, talented shop assistants, and friends with infallible taste who understood you fulfilled this function. In an online shopping experience, however, this experience was rather hard to replicate. Modern relational focused AI recommender systems however have opened an avenue to grow from a convenient storefront to the promise of a lifelong fashion partner committed to making you feel and look your best.

Getting started towards relational online interactions

Since online shoppers are accustomed to the transactional approach, it is important to be explicit upfront about the type of information you require to grow into a trusted assistant within a short time frame and why it is important. This also means you have to offer value beyond the immediate service or product your visitor is there for, which is what builds a deeper relationship and trust.

In fashion, a good example is providing an engaging interactive experience where customers learn something about themselves, such as how their skin undertone affects their ideal colour palette, for instance. Not only does this allow for rich and relevant information to be gathered, but it also builds trust that you as a brand are dedicated to helping your customers be the best version of themselves, beyond any specific product or service you are offering.

Once this initial relevant information position is established, we can start acting on it. Just like in the stylist/trusted friend dynamic, this involves a lot of back and forth where feedback is received and adjusted to narrow down to ultimately the best options.

We also encourage maintaining the transactional approach in conjunction with the relational approach. This is because not all (potential) customers may be willing to engage with the new approach immediately. Transitioning to a relational approach should be seen as an invitation and an enjoyable experience, rather than an obligation.

How to get the maximum out of product recommendations: The summary

To summarise, fashion and apparel companies can gain more benefit and prevent an irrelevant filter bubble by utilising a more personalised, relational recommendation approach that considers customers' psychological demands and purchasing context. Furthermore, to truly understand the underlying needs of customers and evaluate them on more than just their online activity, relevant personal data should be gathered first, ideally in a two-way exchange of feedback between the brand and the customer. Providing fashion e-commerce visitors with the option to share more specific information about themselves, their style, and preferences in an interactive way allows for driving more relevant and personalised recommendations and essentially offering more value in return. In addition, if a more general transactional approach is to be employed, a way out of a possible filter bubble should be considered to prevent the recommendation engine from becoming trapped in the incorrect path. To achieve this, you can provide a few suggested options: include one or two items that don’t fit the current customer profile, request confirmation that the recommended products are relevant, and/or run a ‘dissonance detector’ in the background.

dr. Erik Van Breusegem,
Head of AI at PTTRNS.ai

Are you looking to accelerate your business’s digital transformation? Connect with our digital transformation specialists.