Click-Through-Rate (CTR) is the number of clicks that the item receives divided by the number of times the item is shown.
clicks / impressions = CTR
If we have 5 clicks and 100 impressions, our CTR would be 5%.
Supports
- Google News Personalization scalable online collaborative filteringGoogle News Personalization scalable online collaborative filtering
Problems
Scalability
Item Churn
Baseline
Decayed popularity
Result
Increase in CTR by 38% compared between the proposed method and baseline popularity.
Reference
A. S. Das, M. Datar, ...- Increase in CTR by 38% compared between propose method and baseline popularity.
- A Live Comparison of Methods for Personalized Article Recommendation at ForbesA Live Comparison of Methods for Personalized Article Recommendation at Forbes
Note:
Hybrid of collaborative-filtering and a content-based method that leverage Wikipedia-based concept features
Post-processed by a novel Bayesian remapping technique
Baseline
Popular
Re...- Increase in CTR of 37% (Hybrid with Wiki vs popular)
- Personalized News Recommendation Based on Click Behavior - GooglePersonalized News Recommendation Based on Click Behavior - Google
Bayesian framework
The log analysis reveals that the click distributions of individual users are influenced by the local news trend
Decompose users news interests
...- Improves the CTR upon the existing collaborative method by 30.9%.
- Offline and online evaluation of news recommender systems at swissinfoOffline and online evaluation of news recommender systems at swissinfo
Context-tree recommender systems which profile the users in real-time
the CTR overestimates the actual impact for popular items and thus gives a skewed impression of the actual performance. Th...- Increase in CTR about 35% for longer user sessions (context tree vs random)
- The Youtube video recommendation system 2010The Youtube video recommendation system 2010
Goals: Recent and Fresh, Diverse and Relevant
2 Phase
Candidate generations
[[Co-visitation Recommendation]] (association rule mining)
Ranki...- Increase of over 200% CTR (co-visitation vs most viewed items)
- Recommending Similar Items in Large-scale Online MarketplacesRecommending Similar Items in Large-scale Online Marketplaces
Offline process, generates long-term cluster based on product’s title
e.g.
c1 = {nike, air-max, white, gray, running}
Nike white gray 7 m...- 38.18% increase in CTR
- Post-purchase recommendations in large-scale online marketplacesPost-purchase recommendations in large-scale online marketplaces
Co-purchase mining approach
Results
ebay.com improve 30% in CTR
ebay.de improve 42% in CTR
Reference
J. Katukuri, T. Konik, R. Mukherjee, and S. Kolay, “Post-purchase recommendation...- 30% / 42% increase in CTR
- Optimizing Similar Item Recommendations in a semi-structured marketplace to maximize conversionOptimizing Similar Item Recommendations in a semi-structured marketplace to maximize conversion
Recall + Ranking
Recall: Candidate Generation
Coviewed items
A pair of items that have been frequently viewed together in the same browsi...- Increase of 3% CTR
- A Comparison of Offline Evaluations Online Evaluations and User StudiesA Comparison of Offline Evaluations Online Evaluations and User Studies
Evaluate recommender system
offline evaluation
online evaluation
user studies
Result from offline sometimes contradict from online and user studies
Hu...- Offline is not suitable for evaluation
- CTR better than cite-through rate, link-through rate
- A Comparative analysis of offline and online evaluations and discussions of research paper recommender system evaluationA Comparative analysis of offline and online evaluations and discussions of research paper recommender system evaluation
Offline evaluation contradict with online evaluation
Both CTR and MAP never contradicted each other
It could still be possible that MAP over users will differ
Research Que...- Offline is contradict with online evaluations
- Both CTR and MAP never contradicted each other
- It could still be possible that MAP over users will differ
- TF-IDuF - A Novel Term-Weighting Scheme for User Modeling based on Users Personal Document CollectionsTF-IDuF - A Novel Term-Weighting Scheme for User Modeling based on Users Personal Document Collections
TF-IDuF achieved 5.14% CTR, equal to TF-IDF (5.09% CTR)
Reference
J. Beel, S. Langer, and B. Gipp, “TF-IDuF: A Novel Term-Weighting Scheme for User Modeling based on Users’ Personal Documen...- TF-IDuF achieved 5.14% CTR, equal to TF-IDF (5.09% CTR)
- Evaluating Similarity Measures - A Large-Scale Study in the Orkut Social NetworkEvaluating Similarity Measures - A Large-Scale Study in the Orkut Social Network
6 measures of similarity for recommendations in social network
L1-Norm
L2-Norm
Pointwise Mutual-Information: Positive correlations
Pointwise Mutual-Information: posi...- Use CTR for their measurements
- Recommending ephemeral items at web scaleRecommending ephemeral items at web scale
Generative Clustering Model
Objective
To maximize the total intra-cluster coherence
Use-cases
Naive Bayes for ranking
...- 3-5 folds improvement in CTR and Purchase-through rate
#recommender-system #metrics