Serendipity

Serendipity and RecSys

Scoring functions

Recommendation Candidates

How to evaluate such a system?

Beyond-accuracy measures

• Novelty
• Diversity

Novelty: Unexpectedness

\[\begin{equation*} Unexp(R) = \frac{1}{|R||H|}\sum_{r \in R}\sum_{h \in H} d(r, h) \end{equation*}\] Castells, Hurley, and Vargas (2015)

• distance of recommendations to users’ history
• \(d(\cdot)\) can be any distance measure, we used cosine distance

Diversity: Intra-list diversity

\[\begin{equation*} ILD(R) = \frac{1}{|R|(|R|-1)}\sum_{i\in R}\sum_{j\in R} d(i, j) \end{equation*}\] Castells, Hurley, and Vargas (2015)

• average distance of recommendations to each other
• \(d(\cdot)\) can be any distance measure, we used cosine distance

Diversity: Structural

\[\begin{equation*} DGC(G') = 1 - \frac{1}{|G'|-1}\sum_{i=1}^{|G'|} (2i - |G'| - 1)\frac{deg(w_{i})}{\frac{|E'|}{2}} \end{equation*}\] Sanz-Cruzado, Pepa, and Castells (2018)

• based on Gini index
• degree disparity: how skewed is the degree distribution

DGC: Example

Results

Future Work

• extend knowledge graph with other classes
• extend model to other collections than music
• users’ opinion!

Thanks!