PKP 1

PKP 1 – Who is watching TV at the moment?

Recommendation systems are usually used by companies to tailor product/service offers to each individual customer’s tastes. There are benefits for both. Users find interesting content/products more easily and quickly, and companies find it easier to meet their customers’ needs and increase customer satisfaction, which in turn improves their credibility with the user and their business results.

The project focused on recommending “live content” on TV, i.e. what is currently on air. In the old days, when there were only a few channels available, there were no problems with decision-making. Today, the situation is completely different, there are several hundred channels available and going through them or even looking at the trailer is becoming less and less useful. Our idea was to use a recommendation system to suggest to the current viewer in front of the TV screen some of the most interesting options available to them at the moment (or in the short term, e.g. within an hour). When creating such a recommendation system, it is very important to know who the current user is, as the recommendations are tailored to each individual user. A given TV set may be used by several people. Let’s imagine the following family: a mother, a father, a 12-year-old son and a 5-year-old daughter. Their tastes are different. Since we have to recommend relevant content, our task was to find out who is sitting in front of the TV at the moment. Is it just the son, maybe the daughter, or dad and mum together?

From the recorded data on channel changes, we have built models throughout the project that attempt to distinguish between different users (or their habits) of the same TV channel (same subscription). Patterns in such channel changes, and their repeated occurrence over limited time intervals, can identify the tastes or habits of the user. In our data, the most effective approach has been to label users by the category (genre, TV channel, specific show, etc.) they usually watch and by the number of events they do per time interval. By taking these characteristics into account, we were able to slightly improve the accuracy of the recommendation system. The result of the project is an open source framework that supports all basic operations (including testing recommender systems in several standardised ways) and several recommendation algorithms that take into account different users of the same TV set.