An Overview of ML-powered Recommender Systems
by Aleksander Kijek, Chief Product Officer at Nethone
Machine Learning Algorithms’ potential to revolutionise customer-website interactions
You can have too much of a good thing. That’s what I thought when I was planning my last trip to Berlin. Hotels, hotels everywhere. It doesn’t really matter if you’re looking for the most basic reflection of your definition of a “hotel room”. The sheer number of offers mounting around made me feel claustrophobic. Choice overload, in turn, leads to decision fatigue, which I’ve observed has become a typical affliction of those who shop online. Browsing through thousands of items to find the proverbial needle in the hay stag is downright disheartening, but, to twist the perspective a little, it is equally or maybe even slightly more challenging for online merchants to tailor their offers to their customers’ expectations.
Maybe your view on that last bit will change after you’ve read this short article - at least that’s what I hope for! Essentially, a good customer experience has a different meaning for everybody and so it takes loads of information to understand, truly understand, what an individual’s (not the entire clientele’s) wishes are. Only thus – by understanding a single user – can merchants unburden them of mining through products, programmes, shows, services, profiles or even news, and take the customer experience to a whole different level. The gains are obvious – happy customers equal successful business. So, coming back to my point of departure – it’s not about too much choice. It’s rather about poor filtering. The antidote? Machine Learning (ML) – yours and your customers’ personal adviser.
The overall potential of AI in domains like natural language processing, computer vision and cybersecurity has been vastly explored for some time, but it wasn’t until recently that its capabilities in enhancing recommender systems have been uncovered. Recommendations, as simple as they seem, have a latent potential to maximize profits, minimize risk and revolutionise the way websites interact with users. It’s, after all, the centrepiece of the customer experience offered by Netflix, Amazon, LinkedIn and Google. Today, as an avid Machine Learning evangelist, I’d like to put ML in the spotlight and present an overview of how it can bolster up recommender systems and why it’s a game-changer for e-commerce.
First things first, so let me start off by giving you an intro to recommender systems (RS). There are three types of RS as discussed by Ivens Portugal, Paulo Alencar and Donald Cowan in their paper The Use of Machine Learning Algorithms in Recommender Systems. The three categories have emerged as a result of leveraging selected tools and the flexibility of the modelling of deep learning networks.
The first type, referred to as collaborative filtering, consists in providing recommendations based on what other users of similar tastes and likes have enjoyed. Basically, it means that if a user liked Alien and Interstellar, then probably an avid fan of Interstellar should check out Alien, too. In short, it relies on people’s choices. Content-based filtering, conversely, feeds on (as the name implies) content. It dissects items themselves in search of their distinguishing features. The system will, therefore, advise a Marvel enthusiasts to explore some comic-based productions or other superheroes stories. Finally, hybrid filtering is a form of a trade-off between the two types discussed. In any case, the system accesses user and item data to spot intricacies and similarities on much more profound levels and substantially faster, or in no time in fact, when supported by cutting edge ML-based solutions.
This looks really plain on the surface, yet it offers a massive advantage to customers – often shuttering their preconceived notions by helping them find items that they would never have gone for if it hadn’t been for the recommendation. It’s the science of identifying those nuanced links, threads that are virtually imperceptible to a human, yet evident to a machine. Intuition-driven endeavours to tailor the offerings to what customers are believed to desire are very often daunting and pointless. Going further, employing ML-driven recommender systems can give merchants the power to foresee their needs before they materialize in their minds. A fresh dad won’t be looking for a stroller in perpetuity so perhaps it’s an opportune time to invite him to review your rich portfolio of baby car seats?
Nonetheless, a wonderful sense of what movies users might like or – in business terms - customer empowerment is not the only reason why the empires I’ve mentioned in the beginning tap into ML so willingly. The value that ML brings to businesses is chiefly recognised in the fact that recommender systems study countless data concerning users’ behaviour and transactions. It is all diligently analysed, ready to be applied both instantly as well as in the event of their future interactions with the brand.
In order to make the most of ML in recommender systems, I’d advise innovation-savvy companies to employ A.I.-powered profiling tools, which prove especially successful in this area. Profilers x-ray every individual interaction with the website and collect over 5000 data points featuring users’ software, hardware, network and behaviour. To customers’ delight, they only scan non-personally identifiable information (which is not further administered) to soothe their customers’ privacy-related concerns. The best of them go unnoticed, as unobtrusive as can be, and remain the same for the merchant – no in-house AI talents needed and everything happens in real time. Better still, the data collated can be additionally complemented by the information from databases that merchants create for other, usually analytical purposes. This combination yields profiles that are more accurate and richer than ever!
The depth of information gleaned is indeed unprecedented. The data can easily be converted into business decisions that would never be thought of otherwise. Within the remit of one of our projects that we run as Nethone, we analysed user interactions with a booking platform. What did we learn? For example, we learnt that men were much more dedicated users of the platform, while women, whose lifetime value greatly exceeded that of men’s, tended not to finish their bookings. On this premise, a decision was made to incentivize male users to invite females to join them in enjoying the services. Projected results? Thousands of new female users. We also found out that men, as they’re becoming more and more willing to book wellbeing-related services, made loads of appointments with barbers. Given the increasing value of personal trainers’ services, why not create a joint package right away – get a good haircut and get in shape, too – all nicely priced! And even though this does not necessarily sound like a novelty, it is the rapidity of the system that constitutes an asset of unparalleled impact and brings a new quality into online business.
The mastery of ML-based RS lies in its instantaneousness. What truly matters is the timing – right now, as soon as the user visits the website – regardless of where the they are. The models, if trained skilfully, are able to identify and assign any user to a specific group in no time. Once a customer shows interest in a virtual store, the models can instantly offer them a solution they will find most appealing – even if they weren’t originally looking for it, just like a guy making an appointment to get a new haircut wasn’t going to schedule an evening training session. This opens countless upsell opportunities and so it bears repeating - it all happens in real time. The analyses and subsequent recommendations are being made at a speed not encountered before without compromising their accuracy.
This type of reasoning was only facilitated by relevant data coupled with intelligent software. Such a blend incarnates the new approach that we have devised at Nethone, which is best reflected in the phrase Know Your Users. For us, it denotes a continuous and profound profiling of all users who interact with a website in order to fuel the entire network of ML models so that they can solve diverse problems and improve their performance through mutual reinforcement, regular quality inspections and refining under the supervision of experienced data scientists. In effect, organizations can identify and verify their customers as well as instantly understand them and automatically make decisions on how to cater for their needs – accounting for unconscious ones, too. Technically speaking, to fully enact that vision, it is of vital importance to repurpose the data obtained by existent models so that it can be served as a basis for other analyses. This means that, for instance, a model that predicts the likelihood of a user becoming a frequent shopper can also be capable of identifying users likely to churn within the next 3 months. It is all a rather complex operation and requires the involvement of highly skilled professionals. But is it worth it?
The business results achieved by those who have already implemented such a strategy speak for themselves. The approach is positioned to transform the way merchants interact with their customers via websites eliminating the “unknown” factor. It’s the missing link of how online businesses should interact with their respective clienteles. So why lag behind if you can stand at the forefront of this revolution? With today’s technological advancements, do not be oblivious to opportunities.
Aleksander Kijek, CPO at Nethone ...
|... a global leader in A.I.-powered KYU. Aleksander is a fintech and neuroscience enthusiast responsible for product development, workflow management and ensuring comprehensive operational excellence at the company. Before Nethone, he developed his technical and soft skills as a leader of Polish Integrative System for Alternative Communication and as a coordinator of multiple projects at AJJDC. Profoundly passionate about combining his technological and people skills in the IT world.|