This is the first in the series of three articles where we will be discussing recommender systems. What are their specific challenges? How do we solve the cold-start problem? What is our secret sauce when evaluating these models? What are the specific challenges in domains like telco or e-commerce? Interested? Great, let’s do it!
Sharpen the ax before cutting down the tree
In this part, we will focus on the steps before starting the actual development of a recommender system. We first touch upon the notion of “Personalization” in marketing, and walk further through the steps of creating a business case out of it. We highlight the need to define project goals and metrics, understand the business context, and consider data availability and model explainability when implementing recommender systems. This is easily the most important part, sharpening the ax before cutting the tree, as goes the famous quote by Abraham Lincoln.
You have probably heard it before, personalization is the key, leading to higher sales, clicks, orders, etc. And you are right! Numerous surveys, such as the one from Redpoint or Epsilon, suggest that a modern-day customer expects tailored offers. Well, why wouldn’t he? In the flood of irrelevant offers, he must cut through the internet jungle to find what he needs – ain’t nobody got time for that. Plus, everybody knows “AI” can do just about anything, so why can’t it pick the right color of socks for me?
An email just arrived – WE NEED PERSONALIZED OFFERS! YESTERDAY! Nice, so…where do you start? And how? We call this step a project definition, this is where we define what we want to predict and to whom. Do you want to predict an up-sell to increase spending? X-sell to offer related products? Perhaps acquire-sell to promote new items? Some customers just like their usual cheese in the same quantity, thus simply offering well-known products to keep the customer happy (keep-sell) is the goal. The different goals are usually bundled into one offer, just like when you make a soup – a handful of what always works, an extra spoon of surprise, sprinkled with secret spice.
Understanding the business is, as always, the key. What is more costly? Sending an offer to someone who would not buy it (false positive)? Or not sending an offer to a customer who would otherwise buy it (false negative)? Do we want to predict a yes/no affinity, or a scale of interest?? How does this fit into our data strategy? Where is the business case?
Make sure you know what metric you are optimizing! More clicks? Higher activity? Increased sales? Oh, and how about data? Do you have sufficient uninterrupted user history easily available? This might be a no-brainer for TELCO, but a challenge for websites. We are also asking our customers what kind of model they need. Some want a black-box model – as far as it works, they are happy. Some need more explainability, perhaps because they are obliged by law.
Usage is crucial! Who is going to be targeted? Do you have enough customers with opt-in and valid (email)addresses? Do you need a top-3 recommendation for a user? Or an infinite scrolling list? How is the campaign going to be set up? Our models can have 100% accuracy, but if the campaign is poorly executed, the model won’t save it. Having a good experimental design with A/B testing and comparable groups does not have to be reminded, right…right?
It’s not all rainbows and sunshine. Recommender systems have their own very specific challenges. The one you probably heard about is a cold start – if a new user appears, without knowing his history, it is hard to estimate his preferences. What if a new product appears? Same problem. Different types of information – explicit (buying a product) or implicit (only clicking on a product on a website) have different effects on results.
How to deal with these bumps? Glad you asked! Stay tuned for the second part, where we will dive into more technical details.