Summary: A Serendipity-Oriented Greedy Algorithm for Recommendations

Full text available: proceedings | direct link.

Denis Kotkov, Jari Veijalainen, and Shuaiqiang Wang. 2017. A Serendipity- Oriented Greedy Algorithm for Recommendations. In Proceedings of the 13th International Conference on Web Information Systems and Technologies, Vol. 1. ScitePress, 32–40.

A Serendipity-Oriented Greedy Algorithm

for Recommendations

In this paper, we proposed a serendipity-oriented recommendation algorithm. We had two requirements for the algorithm, while designing it: (1) the algorithm had be easy to apply to any live recommender system and (2) the algorithm had to introduce items with novel features to a user. As a result, we came up with a simple reranking algorithm, which belongs to a family of greedy algorithms.
A general greedy reranking recommendation algorithm have a number of candidate items for reranking. On each step, this algorithm picks a new item to add it to the list of recommendations. To pick a new item, the algorithm uses an objective function, which is recalculated on each step and depends on the items already added to the list.
For example, let's say we have 9 movies to rerank. To initialize the algorithm we pick the first movie. In the following example we picked the one with the highest predicted rating:

We then add movies one by one, based on our objective function. Let's say we want to diversify our list. The score than would depend on how dissimilar a candidate to the movies in our recommendation list:

Movie 4 has the highest score:

We add this movie and recalculate other scores:

Now movie 6 has the highest score. Let's add this movie to the list:

This is a result of the topic diversification algorithm. Our algorithm has a different objective function. Similarly to the Topic Diversification algorithm's objective function, our function takes into account the rating predicted for the user item pair, similarity of the item to items already added to the list, but it also considers similarity of an item to items the user has already rated.