Kotlin Multiplatform & Compose Multiplatform - Starter kit for pragmatics
You Should This! Let me explain to you why.
1. You Should Read This! Let Me Explain You Why
Explaining News Recommendations to Users
R. Blanco, Yahoo! Research Barcelona
D. Ceccarelli, C. Lucchese, R. Perego, F. Silvestri, ISTI – CNR, Pisa
2. Explaining News Recommendations
• News recommendations are shown to users
but
– even in the case that a relevant news item has
been recommended, a user will access it only if
she thinks is relevant before clicking.
• There can be different ways of generating
explanations
– machine learning methods to rank explanations in
order to maximize the usefulness of the
recommendation itself.
3. Explanations Explained
Different Types of Explanations
• Content Based
– SIMILARITY, SIMILARITYSNIPPET,
TARGETSNIPPET, TAGSPLANATION.
• Entity Based
– SHAREDENTITY, TARGETENTITY,
DISTINCTENTITIES, TARGETIMAGE, IMAGES,
SHAREDPLACES, TARGETPLACES,
CATEGORIES.
• Usage Based
– POPULARITY, QUERIES,
TARGETQUERYBIASEDSNIPPET,
SOURCEQUERYBIASEDSNIPPET.
4.
5. Markov Logic Networks
• Rule-based Learning System. Example of rules we used:
– !HasExpl(r,e) => !Relevant(r,e)
• When an explanation e cannot be computed for a pair r, then the
explanation is not relevant
– HasExpl(r,+e) => Relevant(r,+e)
• For each explanation, learn how much it is relevant for a
recommendation
– ShareMainEntity(r) => Relevant(r,+e)
• Learn the relevance of an explanation given the fact that two news
share the same main entity
– TargetHasEntities(r) => Relevant(r,+e)
• Learn the relevance of an explanation given the fact that the target
contains entity annotations
6. Experiments
• Dataset built on a Yahoo! News collection
• Relevance judgments manually built by professional editors
– Goal: understand if the explanation improves appeal of recommended
news
8. Conclusions
• We have defined the problem of news
recommendations explanation
• We propose 16 different types of explanation
– Automatically generated using IR and Entity-based
methods
• We have created a dataset for the learning and
evaluation of explanation techniques.
• We have evaluated a MLN-based method for
approaching the problem solution
– Works better than the static baseline and better than
other state-of-the-art learning methods (e.g. Rank-SVM)