Novelty
It involves the idea if having novel items in the recommendation. Since there are several definitions for novelty (and so metrics), we can classify it into three levels:
- Life level novelty -> a recommended item is novel in the life of the user, i.e. the user has never heard of the item in his/her life. This is the hardest level to evaluate, since we could have to consider information out of the system's context in order to measure what the user knows and do not know.
- System level novelty -> recommended item is unknown for the user according to the user's history consumption. A classic metric was proposed by Nakatsuji et al., which calculates novelty in the recommendation list as the similarity between the items in the recommendation list and in the history of the user ($H_u$), $nov(R_u)= \sum_{i \in R_u} min_{j \in H_u} d(class(i), class(j))$. Another metric considers the sum of the popularity of the items in the recommendation list of the user, where popularity (pop) is the number of users that consumed the item, $nov(R_u)= \sum_{i \in R_u} \dfrac{\log_2pop(i)}{|R_u|}$.
- Recommendation list level novelty -> non redundant items in the recommendation list. No users' information is required in this level. One metris is given by: $nov(R_u)= \dfrac{1}{|R_u|-1} \sum_{j \in R_u} 1-d(i,j)$ where d(i,j) means the distance between items i and j. In this paper the following metric is proposed: $nov(R_u)= \sum_{k=1}^{|R_u|} disc(k)(1-p(seen|i_k))$, where $disc(i_k)$ is a discount of browsing through the list and $p(seen|i_k)$ is the probability of the user has seen the item while browsing.
Tags
#recsys