‍

In the previous episode, we mentioned the model changing events and the concept of trendingness. We suggested that you need to train your systems on detecting model changing events, using the trendingness as a feature. 

The second reason why news is more complicated than other industries is the short shelf life, meaning, the short lifetime of news articles.

‍

Entertainment is relaxing, news is informative and educational.

The way users consume videos or movies is totally different from news. Usually, watching a movie is an evening activity, where you want to relax and take your time. News doesn’t get that tranquility since you usually consume them “in-between two things”, when you have a 15 minutes break or while commuting. With news, recommendations are time-poor. They need to provide the same relevance but with lower time and digital space to convince the user. 

For this reason, real-time computing is a crucial element for impactful recommendations. Given that, on the one hand, the world is changing pretty quickly, and on the other hand, your interests at different moments of the day may be different, re-ranking your recommendation is crucial. To achieve this capability, you need strong mathematical skills to reduce computational complexity while not losing any accuracy. 

‍

Re-ranking recommendations are the key in real-time personalization.

From our experience, retraining our models every 30 minutes and re-ranking the recommendations in real-time is a very useful method. This way, the user gets to see relevant information specifically on the topic that showed previous interest in. But watch out, because abusing this would lock your readers in an echo chamber. In the next episode we are talking about the filter Bubble and how to avoid it. Stay tuned.