# Serendipity --- #### A survey of serendipity in recommender systems > KOTKOV, Denis; WANG, Shuaiqiang; VEIJALAINEN, Jari. A survey of serendipity in recommender systems. Knowledge-Based Systems, 2016, 111: 180-192. --- ## Content * Introduction * Concepts * Approaches * Evaluation strategies * Future directions * Conclusion --- ## Introduction --- * To overcome the flood of information, online stores have widely adopted recommender systems * Most recommendation algorithms are evaluated based on accuracy will cause over specialization problem --- * One of the ways to overcome the overspecialization problem is to increase the serendipity of an RS --- * RQ1 What is serendipity in recommender systems? What makes certain items serendipitous for particular users? * RQ2 What are the state-of-the-art recommendation approaches that suggest serendipitous items? --- * RQ3 How can we assess serendipity in recommender systems? * RQ4 What are the future directions of serendipity in RSs? --- ## Concepts --- ### Components of serendipity * Serendipity is a complex concept, such as relevance, novelty and unexpectedness * we might regard an item as relevant for a user if she/he has gaven it a high rating and/or if she/he has purchased it --- ### Novelty 1. Item novel to an RS 2. Forgotten item 3. Unknown item 4. Unrated item --- ### Serendipity in a computational context ==**Pure**== serendipity is not amenable to generation by a computer. --- * the concept of serendipity includes four components 1. Prepared mind 2. Serendipity trigger 3. Bridge 4. Result ---  Fig. 1. Euler diagram of items from a user’s point of view at a given moment of time (snapshot). --- ## Approach --- ### Recommendation algorithms * Content-based filtering (CBF) * Collaborative filtering (CF) * Hybrid filtering --- ### Serendipity-oriented approaches * Reranking algorithms (Reranking) * Serendipity-oriented modification (Modification) * Novel algorithms (New) --- ### Serendipity-oriented recommendation algorithms * Pre-filtering * Modeling * Post-filtering ---  --- ## Evaluation strategies --- * offline evaluation * online evaluation --- * component metrics * full metrics --- ### Component metrics * Novelty * Unexpectedness --- #### Novelty $$ n o v_{v a}^{d i s t 1}(i, u)=\min _{j \in I_{u}} \operatorname{dist}(i, j) $$ $$ n o v_{v a}^{d i s t 2}(i, u)=\frac{1}{\left|I_{u}\right|} \sum_{j \in l_{u}} \operatorname{dist}(i, j) $$ $$ \operatorname{dist}(i, j)=1-\operatorname{sim}(i, j) $$ --- ### Unexpectedness $$ P M I(i, j)=-\log _{2} \frac{p(i, j)}{p(i) p(j)} / \log _{2} p(i, j) $$ --- $$ \text { unexp }_{k a m}^{c o-o c c 1}(i, u)=\max _{j \in L_{\omega}} P M I(i, j) $$ $$ \operatorname{unexp}_{k a m}^{c o-o c c 2}(i, u)=\frac{1}{\left|I_{L}\right|} \sum_{j \in l_{u}} P M I(i, j) $$ --- ### Full metrics $$ \operatorname{ser}_{m u r}(u)=\sum_{i \in R_{a}} \max \left(\operatorname{Pr}_{u}(i)-\operatorname{Prim}_{u}(i), 0\right) \cdot \operatorname{rel}_{u}(i) $$ --- ## Future directions * Popularity and similarity in RSs * Context-aware RSs * Cross-domain RSs * Group RSs --- ## Conclusion * RQ1 What is serendipity in recommender systems? What makes certain items serendipitous for particular users? * RQ2 What are the state-of-the-art recommendation approaches that suggest serendipitous items? --- * RQ3 How can we assess serendipity in recommender systems? * RQ4 What are the future directions of serendipity in RSs? --- ###### tags: `paper`