Tagommenders: Connecting Users to Items through Tags
Written by Shilad Sen, Jesse Vig, and John Riedl (2009)
Presented by Ken Hu and Hassan Hattab
Overview
• Introduction• Tagommender
o Philosophyo Dataset
• Tag Preference Inferenceo Approacho Methods
• Recommenderso Implicito Explicit
• Results• Conclusion
First, Recommenders.
• What is Recommender system?• Two Main tasks
o Recommend.o Predict.
Recommender Systems
• Types of recommender systems:o User-based: decides according to the user's
previous choiceso Item-based: decides according to related items
to a selected itemo SVD
• Problem: These methods don't consider the content of the item.
• Solution: Content-based Recommenders
Overview
• Introduction• Tagommender
o Philosophyo Dataset
• Tag Preference Inferenceo Approacho Methods
• Recommenderso Implicito Explicit
• Results• Conclusion
Tagging Systems
• Uses tags to address (categorize) items to users• Tags are created by general users (More
meaningful )
Tagommenders:
• Basically, they combine Recommenders (content-based) and tagging systems.
• Two main parts for Tagommenders:o They infer users’ preferences for tags based on
their interactions with tags and movieso and they infer users’ preferences for movies
based on their preferences for tags.
Tagommender's data set
These are collected from the MovieLens website.
• Movie Rating • Movie clicks• Tag applications• Tag Searches • Tag Preference Ratings
Tagommender's data set
Overview
• Introduction• Tagommender
o Philosophyo Dataset
• Tag Preference Inferenceo Approacho Methods
• Recommenderso Implicito Explicit
• Results• Conclusion
Tagommender's Cycle
Inferring Tag Preference
• Inferring Preference using Tag Signals (Direct)
Inferring Tag Preference
• Inferring Preference using Item Signals (indirect)
Inferring Preference using Item Signals
• Sigmoid transformation is used to calculate the weight of movie m to tag t
Inferring Preference using Item SignalsMethods1.Movie-Clicks2.Movie-log-odds-clicks3.Movie-r-Clicks4.Movie-r-log-odds-clicks5.Movie-Rating6.Movie Bayes
1- Movie-Clicks:
set of movies clicked by user u
2- Movie-log-odds-clicks
3- Movie-r-Clicks4- Movie-r-log-odds-clicks
• The only difference is Movie-rating is counted rather than movie clicks
5- Movie-Rating
• A user’s preference for a tag is the average rating for a movie under that tag.
user u's rating for movie m
6- Movie-bayes
• A bayesian generative model for users rating for a certain tag.
• if the tag is relevant to a rating then the rating will be chosen from the user-tag-specific distribution
• Else, it will be chosen from the user background rating distribution
Which one is better?
Overview
• Introduction• Tagommender
o Philosophyo Dataset
• Tag Preference Inferenceo Approacho Methods
• Recommenderso Implicito Explicit
• Results• Conclusion
Recommenders
• Implicito Tag data onlyo Recommend onlyo 2 algorithms
Implicit-tag Implicit-tag-pop
• Explicit Algorithmso Use users' movie ratingso Recommend and predicto 3 algorithms
Cosine-tag Linear-tag Regress-tag
Implicit : Implicit-tag
• Vector Space Modelo Inferred tag preferenceo Relevance weight
Implicit : Implicit-tag-pop
• Implicit-tag with movie popularity
o Tag > clicks, clicker count > click counto Linear estimation of log function
Recommenders
• Implicito Tag data onlyo Recommend onlyo 2 algorithms
Implicit-tag Implicit-tag-pop
• Explicit Algorithmso Use users' movie ratingso Recommend and predicto 3 algorithms
Cosine-tag Linear-tag Regress-tag
Explicit : Cosine-tag
• Cosine similarity: rating vs tag preference
Explicit : Linear-tag
• Least-square fit linear regression
Explicit : Regress-tag
• Linear-tag with similarity between tags • SVM was best to estimate h
o Robustness against overfitting
Overview
• Introduction• Tagommender
o Philosophyo Dataset
• Tag Preference Inferenceo Approacho Methods
• Recommenderso Implicito Explicit
• Results• Conclusion
Results : Background
• Comparisonso Top-5
Compare top five recommendations
o MAE Average error of
prediction
• Competitorso Overall-avg o User-avgo User-movie-avgo Explicit-itemo Implicit-itemo Funk-svdo Hybrid
Regress-tag + funk-svd
Results : Top-5
Results : MAE
Overview
• Introduction• Tagommender
o Philosophyo Dataset
• Tag Preference Inferenceo Approacho Methods
• Recommenderso Implicito Explicit
• Results• Conclusion
Conclusion
• Introduced recommender algorithms based on user suggested tags (Tagommenders)
• Best at recommendation tasks• Adds value at prediction tasks
o Hybrid predictors does very well• Other advantages
o Ease to explaino Algorithmic evaluation of tag quality
Questions?