In the emerging business need to stay competitive, increase conversion rates, boost loyalty programs and/or improve customer experience during shopping – recommenders have become an integral part of data-driven strategies.
Depending on the end goal, available data and technical resources – there are different types of recommender models that can be of use. The final output is to come up with a proper set of recommendations to propose/present to the given customer/visitor. Let’s start from the basics – non-personalized and personalized recommenders.
Non-personalized recommenders, as the name suggests, provide recommendations which in most cases are the result of generic patterns identified within the aggregated data. At Things Solver, we refer to them as starter-pack recommenders. Example given, recommendations based on all time best-seller, or recently trending products are the most popular non-personalized recommenders. They could work in a good amount of cases, but do not really show that we listen and understand our customers. An additional type of non-personalized recommenders are the new product recommendations, which are related to the cold-start problem where most of the models cannot address the need to recommend something new as relevant to the customers. There are methods for overcoming these, but they require using intelligence and analytics in order to find a workaround (like relying on the recommender pattern of most similar products based on which the new products can be offered, or finding early-adopters who are most likely to be interested in new products).
Starter pack recommenders are cheap, simple and convenient for applications where we don’t know who the customer is, or do not identify and track customer purchases over time, so the aggregated statistics is the only way to go – cash desk recommendations, promo banners in stores/streets, homepage recommendations on webshop,… However, if this information is available, non-personalized recommenders can be extremely outperformed by personalized recommenders.
Introducing the concept of “personalization”
As the data on disposal increases, so does the potential for developing a proper recommender model also increase. Business has to be aware of this – it is a great shame not to utilize the available data to get the most of it. Personalized recommenders are based on patterns identified within the data. The simplest, yet pretty powerful, recommender is history-based recommender – recommending products a customer has already purchased before and has a tendency to repeat these purchases. Is it personalized – yes, since every customer will get a different set of products recommended. Is it relevant – yes, since there is a tendency towards purchasing the products identified in historical transactions. Is it AI – not necessarily, since it can be developed through a couple of filters, groupings and aggregations. Another type is intent-based recommender – proposing products a customer has viewed several times on a web shop, or added it to cart – but not purchased it. It is personalized, it is relevant, and it is a result of a simple combination of filters and aggregations.
Personalized recommenders are a must-have when there is any kind of customer identification available and history about interactions/purchases of products collected. They are great for the so-called exploitation strategy – proposing relevant products based on the interest. But what about the other interests the customer has not yet explicitly shown? This is one of the areas where AI may come in handy – it can help us overcome this and reveal customer intentions even before they have been shown or even developed by the customer himself. How? By analyzing other customers who are similar, or analyzing other products which are somehow associated with products of his interest.
Flavoring the personalization with AI
AI recommenders are an effective way to analyze customer preferences, intentions and needs at every single occasion. They are learning from the data – and constantly improving, as the available data increases. When they are aided by a feedback-loop, either through some kind of defined rules and filters, or by reinforcement learning, they are the most powerful tool for tailor-made personalization. There is a wide range of recommenders developed so far, using the AI mechanism in different prisms, from simplest ones like collaborative filtering, to the more complex ones – like content-based recommenders based on NLP algorithms, or recommenders based on graph neural networks. AI recommenders are great for identifying long-term patterns which reflect customer preferences with recent events, i.e. identifying that a customer is tech savvy, having preferences towards black-color gadgets, thus recommending him a black iWatch of newest generation while he’s scrolling through Apple products.
AI recommenders are effective and fully-personalized, but they could cause problems with resources and speed, and different models often depend on different dimensions. Some models highly depend on the number of products or transactions, whereas others depend on the available product attributes – when their performances induce a headlong dash into memory and CPU usage, and eternity to output results. This is why it is important to adjust technology used to the specific business need and resources being available. Cloud-based, maintained services and serverless applications, followed by large-scale data processing frameworks like Spark could be life-savers, as we have learned throughout our journey to developing top-performing recommenders.
Incorporating recommenders into the business model
Tailor-made recommenders are a top class ticket to a successful marketing/sales strategy. However, having the relevant recommendations does not always ensure results. The story is getting complicated once the proper timing and channel are included in the equation. A customer has different needs, with different importances on different occasions. Thus, what is the most important is to develop a recommender system – which wraps up a recommendation model and upgrades it with additional insights which provide the answer to the following questions:
- What to recommend?
- Where should the recommendation be placed in – which channel?
- When is the right time to trigger a recommendation – during the browsing process, on Friday evening, or outside of working hours?
- Who is the right audience for a specific offer that should be placed?
Throughout our recommender system journey we have learned that we should start from the beginning and cover the most basic business needs, and then add complexity and flexibility as the business needs grow. We have clients of different sizes, thus we needed to adapt to their way of working. That is why we have developed different recommender types – starter-pack recommenders, intelligent personalized recommenders, and AI recommenders. We support both offline and online recommendations. We also developed a mechanism to identify the right timing for recommendations. 30% conversion rate increase and 20% higher average basket value are metrics we proudly share with our clients. Results obtained so far are giving us strength and motivation to continue developing new stuff, like hybrid recommender systems combining different recommender models and reinforcement learning recommenders. Combining recommendations with search history, in order to get efficient product ranking. The goal is always the same – enable the best possible shopping experience to every customer. Thank you for reading!