Skip to content

The Cold Start Problem and the Power of Real Time Analytics

The Rise of Recommender Engines

If you regularly browse through the internet, chances are you benefit from recommender engines every day, as they are increasingly becoming the standard not only for e-commerce, but anyplace where personalization improves user experiences in the virtual world. While Amazon helps you to find the items you need, Netflix knows which movies you like to discover, and YouTube knows which video you are craving to watch. While it is fair to say that those companies have more than average resources to develop such engines in-house, luckily, there is a strengthening development of democratizing such technologies and spreading them across broader landscapes, enabling smaller players to profit from them as well.
A growing number of players is helping to deploy such technologies by enabling facilitated integration to existing systems. However, there are significant differences in the quality of the engines, the approaches that are taken, ranging from self-serving to high involvement, and the state of technology used, especially regarding the use of machine learning to improve recommendations. To better determine which approach is the best fit for specific cases, it is extremely helpful to understand more about how recommender engines work and what kind of problems they must solve. That’s why today, we take a closer look at a notorious problem in the space of recommendation engines as well as one of the most promising approaches to solve the issue: The cold start problem and the power of real time analytics.


The Cold Start Problem

Recommendation engines often use historical data to generate recommendations. The historical data includes for example information about what the user bought in the past, feedbacks the user provided, or videos the user watched multiple times. Based on this information the recommender engine builds up an understanding of the user’s preferences and is able to generate recommendations for items the user is likely to engage with. Such historical data can, however, be difficult to obtain. Take for example a customer, who is not logged in and visits a new website for the first time. As there is no historical data to rely on, it becomes impossible to provide user-specific recommendation based on past actions. This scenario is referred to as the cold start problem. 

There are different approaches to tackle the cold start problem. A popular option is to start with popularity-based recommendations. This means, whenever a new user visits the website, items which are popular amongst other users at the moment are recommended. So, if there is a new trending season of Stranger Things and you visit Netflix for the first time, chances are Stranger Things will be recommended to you. While this approach is already an improvement to a static approach, the recommendations are not personalized and therefore do not at all take into account the diverse preferences and tastes of users. Depending on the use case, popularity-based recommendations can increase customer engagement slightly to significantly, depending on the diversity of the customer base. 

Another way to approach the problem is by actively asking new users about feedback. By answering questions such as “in which subjects are you interested in?”, “what are your favorite genres?” or “about what do you want to learn more?” the user gives valuable feedback and allows for a personalized experience. However, depending on the situation, going with active feedback questionnaires is not the right option as users get annoyed or are just too lazy to answer. Luckily, meta data can help out to some extent by providing little information without explicit feedback. This information can for example include the location, the applied search machine, or the type of the device being used. If someone is for example using the newest iPhone in contrast to an IBM machine, first assumptions can be made to help generating more specific recommendations. While these mentioned approaches present a first step to tackling the cold start problem, we are going to take a closer to look at the most powerful solution yet.  


The Power of Real Time Analytics

The most powerful approach to the cold start problem is by analyzing real time data. Using a machine learning recommender system with real time analytics enables to generate personalized recommendations the minute a user starts browsing on a website or application. As soon as the user interacts with the site, the system starts collecting behavior data. With continuous interactions the recommendations become increasingly accurate and personalized. By constantly collecting massive amounts of data for each user, real time analytics make it possible to create real one on one experiences and thus lift user experiences to new levels.

The benefits of real time analytics go way beyond to just mitigating the cold start problem. Preferences can change in an instance, making it essential that recommendations stay up to date and evolve just as the user does. In some scenarios, user behavior changes drastically. Imagine someone visits an e-commerce shop in desperate search for a present to their new in-laws. Relying on historical or meta data will probably deliver results of little interest to the user in that moment. However, using machine learning models fed with real time data enable recommenders to be agile enough to come up with helpful recommendations instantly, making the user’s search experience more convenient, and hopefully more successful.

At Flike we develop machine learning recommender systems with real time analytics that can be integrated to your application within minutes. If you want to learn more about real time analytics and our services, reach out to us at

Blog comments