Recommendation algorithms

The different recommendation algorithms

Even though there are many different algorithms for computation, these are ultimately based either on the approach of comparing the items to be recommended with each other (editor's note: In this context, the term "item" generally describes any object to be recommended, such as CDs, videos, news articles, books or vacuum cleaners) and to calculate similarities or to analyze usage by the users, thereby identifying which users have things in common in their usage profile. The first approach is known as "content-based recommendation" and the second as "collaborative recommendation."

Content-based recommendation

For all content-based recommendation processes, the metadata that characterize an item are of key importance. If the metadata are missing, e.g. if the artist or music genre is not indicated for a piece of music, it is impossible to compare the items and to obtain information pertaining to the items' similarity. If metadata are available, content-based algorithms typically provide very good results. In principle, the problem is the same as for a search: Only if sufficient structured knowledge about the content is available can good recommendations be made. For some content such as news articles, websites or other text-based content, such indexes can be generated by machine, but for films or images this cannot be done in sufficient quality yet. Whenever few metadata are available or if they are incomplete, recommendations based on the process of collaborative recommendation can still provide good results.

Collaborative Recommendation

Collaborative recommendations are based on the understanding that the "wisdom of the crowd" often leads to better results than elaborate calculation or data analysis. For all collaborative processes, historic usage data are indispensable. This makes it possible to determine independent of the items' metadata or across various product groups with different metadata which users tend to consume the same content. In that case, recommendations can be made based on those items that a similar user has already consumed. The favored calculation of popular items is one variation of the collaborative recommendation. However, all collaborative processes have one weakness: the so-called cold start problem. If new items are added, they have not yet been consumed by any users and can therefore not be recommended to any other user either. Furthermore, a user who registers for such a system for the first time will also experience a cold start problem. Due to the non-existing history, he or she has no similar users and, as a consequence, no targeted recommendations can be generated via this route.