Subscribe to our blog

How Behavior-Based Product Recommendations Warm Up The “Cold Start” Issue

May 09, 2019
5 min reading time

Product Managers in the eCommerce space often are tasked with solving this problem: how do you guide a customer’s journey if you don’t know their past history or experience with your brand? More specifically, how do you create product recommendations that are personalized to your users, especially new users? Maybe your customers don’t typically login, they clear their cookies or login infrequently, or they haven’t been on the site before. How do you figure out what types of products are related to a new item without any historical data? Oftentimes called the “Cold Start” problem, a computing issue can arise when there isn’t enough data to make inferences about the user or the piece of content, or how the user would interact with that content.

A quick Wikipedia search presents a couple of uses for the term: “Cold Start” in computing and “Cold Start” in automotive. “Cold Start” when referring to a vehicle perhaps even sheds light on the issue at hand, as the automotive definition is the problem that can arise when starting a car engine in cold temperatures. Essentially, an engine has more difficulty igniting when the temperature is cold, but with modern technology, such as electric starters, manufacturers can greatly increase the dependability of ignition.

Similarly, your product recommendation “engine”, once ignited with proper personalization, will drive your customer journey and success in conversion. But in a “Cold Start” setting for your site, where you have sparse data because of a new user, an infrequent user, or a new item for sale, your product recommendations will have difficulty propelling the customer journey forward and will be challenged to convert that customer without the use of modern technologies. If you are relying on more antiquated approaches to personalizing your product recommendations, such as a rules-based infrastructure, customer logins and/or purchase history alone, the “Cold Start” problem remains unsolved. It is only through the more modern approach of real-time machine learning that you can quickly choose the most effective path for conversion. Think of real-time machine learning tools (like LiftIgniter) as being the electric starter for your personalization engine.

Think of real-time machine learning tools (like LiftIgniter) as being the electric starter for your personalization engine.

There are four situations on an ecommerce site where “cold starts” are particularly common:

  1. New Users
  2. Infrequent, Sporadic Users
  3. Not-Logged-In Users
  4. New Products in your Catalog

New Users:

New users are a blank slate in the world of personalization. Net new users provide a challenge even when using real-time behavior-based personalization, as there are no obvious indicators or signals to use for predicting what the customer is looking for yet. That being said, the recommendations can still be personalized using trending items related to the page the users are on, what time of day it is, what device they are using, and other signals from the very first moment.

With personalization solutions using behavior for predictions, the “Cold Start” problem goes away extremely quickly. After as few as 3-4 behavioral events from a homepage, such as clicks or spending a certain amount of time on page, real-time machine learning already has enough specific information to have a fairly concrete prediction of what the user is looking for. Also, because users often times being driven to a product page directly through marketing efforts, companies who use real-time behavioral analysis, like LiftIgniter, can immediately make a fairly good guess at what the user might click on next right from the very first page. Conversion improves by providing timely personalization – in other words showing the right recommendations to the user at that moment in time. The engine is ignited by what that user is doing right then, not based on personal information which merely tells you who the user is.

Infrequent Users:

Infrequent users can pose a similar problem – for instance if you are selling something that is bought only seasonally like a cruise package, or have higher price point items that are sold every once in a while, such as a luxury item like a high-end watch or handbag. In this case, using behaviorally-based indicators to determine how to shape your customer journey is even more essential, as you don’t have much opportunity to get it right. You’ve spent so much time and so many marketing dollars getting your users to your site, and if you aren’t optimizing your customer’s discovery in real-time on every click, you are leaving money on the table.

Infrequent users can also still be loyal customers, perhaps even typing in your homepage instead of being driven to a product page, which present the “Cold Start” problem again. In such a case, personalize the homepage by using what little you do know about the user and still showing real-time trending items, and new items similar to past purchase behavior for example. Presenting the user with a few different varieties of product recommendations on the homepage can help warm up the user with more timely options which then in turn help inform the customer’s real-time preferences for the subsequent customer journey.

A note on complexity:

While far more effective on beating the cold start issue, arriving at the right product recommendations using real-time behavior and a variety of personalized options right from the start is computationally complex. Taking into account each click, each click path, what options are not clicked on in the product recommendations, the amount of time the user takes on each path, plus a multitude of traits about the user and device, really does require the use of real-time machine learning to optimize your customer journey.

No alt text provided for this image

As an example, let’s say you drive 100,000 new prospective customers to the same product page with a “you may also like” section that has 5 additional products. If each of those 100,000 customers clicked the product recommendation section on the second page, there are already 25 different paths they could have taken (5 product recommendations x 5 product recommendations + 5 top recommendations). If you were to create a chart just to keep track of which of those 30 different paths the customer took, you would already have a chart with 3 million cells. In this example, you aren’t even taking into account which paths weren’t taken, or who converted/who didn’t convert.  In order to truly leverage the behavioral information and create the absolute best predictive recommendations for conversion, you have to use something powerful and constantly learning to produce results.

Not-Logged-In Users:

As mentioned in my article on top 5 considerations when developing your recommendation strategy, forcing the user to log in before they start browsing will kill your business. I’ve seen quoted numbers ranging anywhere from 23% - 35% of cart abandonment occurs when the online store forces the user to login before purchase, so it is far better strategically to drive engagement and conversion of new or non-logged in users through personalization. Furthermore, standard practice suggests eCommerce companies provide a guest checkout so that the user doesn’t feel forced into divulging personal information.

Even if the user isn’t logged in, your recommendation solution can still maintain an anonymous 1st party cookie to aggregate the user’s behavior. Personalization can take into account the user’s previous visits, and can use that data to further enrich the recommendation sets.

New Items/Products for Sale:

New products that have been added to your catalog can also cause a similarly challenging “Cold Start” problem computationally. When items are added to your online catalog, prediction sets can start suggesting these new items to explore what statistically tie them together to other items or behaviors. Until this new item has had enough time to be tested out, the personalization can use metadata to link this item to similar items. Even without explicit metadata, the engine can also start inferring relationships using other AI techniques like Natural Language or Image Processing. Another common technique is to apply a boost for that product within the product recommendation area to artificially promote the product. The % of boost can even decay over time, once the item has had enough time to build up interactions and relationships to other products from user activities.

It's almost summer time - don’t be stuck in the cold without an electric starter

In conclusion, much like the electric starter improved ignition in car engines, using the modern technology of real-time machine learning personalization can improve ignition in your personalized product recommendation engine. Leverage massive amounts of data such as user behavior, user data, metadata and infer relationships based on other AI techniques to optimize your customer journey for conversion. Without requiring the use of personal data, logins, or complex database integrations, warm up the “Cold Start” problem for your users, and create a better experience that converts a lot more.

Ready To Learn More?

Schedule a 15 minute demo today and see how LiftIgniter can increase conversions, improve subscription rate and drive user engagement for your site without building new infrastructure.

Request Demo