ndey96
commented
5 years ago - diversification has become a leading topic because it can solve the over-fitting problem and improve user experience
- all RSs only consider items that have not been consumed by the user
- user feedback collection
- explicit: ask user to provide rating
- implicitly: track user’s activity
- general RS research challenges
- data sparsity: users wont rate every item in the catalog
- cold start: hard to deal with new users or content
- big data: limits types of algorithms that can be used
- over-fitting/overspecialization: keep recommending items that the user likes, but not capturing the user’s full preferences
- diversity definitions
- average dissimilarity between all pairs of items in the result set (CURRENTLY PREFERRED)
- Gini coefficient - measure of distributional inequality
- diversity between two items is based on item relevance, similarity, and places in the ranked list
- user perceived diversity questionnaire
- diversity part of nDCG (net discounted cumulative gain) measure
- diversity presented as a nDCG measure
- combination of genre coverage (how many genres present in ranked list) and non-redundancy (genres do not repeat on the list) -> (shows promise to be improved with more metadata)
- evaluation methods
- measure diversity, similarity, and relative gain on 7 RSs on a 1000 job advert dataset
- simulated environment (gini)
- measured diversity on MovieLens 1M dataset with several algos
- 20 participants answered a questionnaire about diversity of recommendations
- calculated diversity on MovieLens and last.fm dataset
- last.fm
- movielens and netflix
- 250 volunteers evaluated the diversity of presented recommendations using a 7 point likert scale
- diversity algorithms
- re-ranking recommendation list using: item popularity, reverse predicted rating value, item average rating, item absolute likeability, item relative likeability
- change distribution of rec. items in terms of popularity
- with 1% precision loss, increased percentage of long tail recs from 16% to 32%
- maximization of parameterized combined objective function representing accuracy/diversity tradeoff, using greedy relaxation and quantization algorithms
- increased diversity, measured by evaluation of precision and recall against the novelty of recommendations
- diversity determinant is genre difference among films
- evaluated algos with users
- users noticed high-diversity items, found them interesting
- ClusDiv: items are clustered rec. list built by selecting items from different clusters; aims to maximize diversity levels on the rec. lists
- increased diversity with no impact on accuracy (z-diversity and accuracy using recall)
- SM - probabilistic specification maximizer model
- increased diversity measured as inter-list diversity
- Pareto-efficient multi-objective ranking and rec. list build
- increased diversity: diversity measured by distance, accuracy measured by precision and recall
- re-ranking recommendation list using: item popularity, reverse predicted rating value, item average rating, item absolute likeability, item relative likeability
- experiment showed that participant personality did not affect recommendations diversity
- experiment showed that diversity positively impacted user satisfaction
https://www.sciencedirect.com/science/article/abs/pii/S0950705117300680