IPTV Recommender Systems Paolo Cremonesi
Agenda 2 IPTV architecture Recommender algorithms Evaluation of different algorithms Multi-model systems
Valentino Rossi 3
IPTV architecture 4 Live TV Set-top-box (decoder) VOD Content Provider Service Provider Network Provider Customers
IPTV architecture 5 IPTV is a video service supplied by a telecom service provider that owns the network infrastructure and controls content distribution over the broadband network for reliable delivery to the consumer (generally to the TV/IP STB). Services Broadcast TV (BTV) services which consist in the simultaneous reception by the users of a traditional TV channel, Free-to-air or Pay TV. BTV services are usually implemented using IP multicast protocols. Video On Demand (VOD) services, which consist in viewing multimedia contents made available by the Service Provider, upon request. VOD services are usually implemented using IP unicast protocols.
IPTV Platform: Now 6 HUNDREDS LIVE CHANNELS CUSTOMERS FACE DIFFICULTIES FINDING THE RIGHT CONTENT THOUSANDS VOD ITEMS CUSTOMER PURCHASES CUSTOMER FRUSTRATION
IPTV Platform: with a recommender systems 7 From this. Today recommendations, based on your personal taste, are: To this.
IPTV recommender needs 8 Improve user satisfaction Sell new content to users VOD Pay-per-view channels Targeting advertisement
Agenda 9 IPTV architecture Recommender algorithms Evaluation of different algorithms Multi-model systems
Recommender System: how it works 10 USER DATA USER S TASTE FRUTIONS AND RATINGS CONTENT METADATA RECOMMENDER SYSTEM CONTENT RECOMMENDATIONS
Problem formulation 11 Users ratings Recommender Items metadata Ranked list Item1 Item2 Item3... ItemX Top N
Recommendation techniques 12 Recommender algorithms Collaborative Filtering User based Item based Content-based Filtering Users with similar taste Similar Items
Memory vs. model based 13 Memory based Model based User-based X Item-based X Dimensionalreduction X Content-based X
Collaborative Filtering 14 4 5?3 User-based similar users rate an item similarly Item-based similar items are rated by a user similarly 2 2 NB: similarity means correlation Neighborhood
Collaborative filtering: User Rating Matrix 15 User Item
User rating matrix URM 16 I1 I2 I3 I4 I1 I2 I3 I4 U1 3 4 0 1 U2 2 2 1 0 U3 2 0 0 4 U4 1 5 0 1 U5 3 0 1 0 Explicit URM U1 0 1 0 1 U2 0 0 1 1 U3 1 1 0 0 U4 1 1 0 1 U5 0 1 1 1 Implicit URM
Dimensional-reduction collaborative model 17 items and users can be described by a number (K) of unknown features a uf : describes if feature f is important for user u b if : describes if feature f is present in item i r ui : rating assigned by (or estimated) user u to item i r ui = k f=1 a uf b if
Singular Value Decomposition 18 m x n A R = U S V T m x n m x k k x k k x n S k RA = k U k V k T
Singular Value Decomposition 19 R = U S V T V V T = I U T U = I
Singular Value Decomposition 20 R k = U k S k V T k V k V T k = I U T k U k = I R k : best rank-k approximation of R according to the Frobenious norm not according least square error!!
Folding-in 21 New rows/columns of A are projected (folded-in) in the existing latent space without computing a new SVD e.g., a new user u u = uv k S k -1 S k V k A k U k u u
Collaborative Filtering: pro & cons 22 Pro: There is no need for content Cons: Cold Start: we needs to have enough users in the system to find a match. Sparsity: when the user/ratings matrix is sparse it is hard to find a neighbourhood. First Rater: cannot recommend an item that has not been previously rated anyone else Popularity Bias: cannot recommend items to someone with unique tastes. Tends to recommend popular items (dataset coverage)
Content-based Filtering 23...mettendo a punto una scoperta che potrebbe portare al primo uso terapeutico della controversa procedura. Se gli studi animali si riveleranno promettenti, i ricercatori potrebbero cominciare a mettere alla prova le nuove cellule su occhi umani da qui a due anni... Term 2 Term 3 Term 1 Similar items contain the same terms The more a term occurs in an item, the more representative it is The more a term occurs in the collection, the less representative it is (i.e. it is less important in order to distinguish a specific item)
Content-based filtering: Item-Content Matrix 24 Word Item
Content-based Filtering: techniques 25 User-item similarity Term 3 Term 2 Term 1
Content-based Filtering: pro & cons 26 Pro: No need for data on other users No cold-start or sparsity problems, neither first-rater Able to recommend to users with unique tastes Able to recommend new and unpopular items Can provide explanations about recommended items Well-known technology Cons: Requires a structured content Lower accuracy Users tastes must be represented as a function of the content Unable to exploit quality judgments of other users
Content-based Filtering: Latent Semantic Analysis 27 V A A k = U k k m x n svd S k T -Terms in rows -Items in columns U k * sqrt (S k ) pseudo terms V k * sqrt (S k ) pseudo items cosine A k
Recommender architecture 28 Resources management Items Storage Features extraction Features representation Filter Users management Compute useritem correlation Items retrieval Items recommendation Users Infer and learn profile Interests/tastes representation feedback Explicit vs implicit ratings
Datasets 29 Real datasets composed by movies and user fruitions, plus some extra information User-item rating matrix 23942 users 564 movies 56686 ratings Movie Meta-data (textual information) Title Genre Director Cast Duration
30 Implicit vs Explicit Come determinare il rating implicito VOD TV (EPG)
Some problems with IPTV recommender 31 Cold start Multi-language content (e.g., Switzerland) New user problem (user-based algorithms) New item problem (all collaborative algorithms) Semantic problem (e.g., house and home)
Agenda 32 IPTV architecture Recommender algorithms Evaluation of different algorithms Multi-model systems
Problem 33 Many works do not describe clearly the methods used for performance evaluation and model comparison Different dataset partition methodology and evaluation metrics lead to divergent results The Netflix prize has improperly focused the research attention on Hold-out RMSE
Objective 34 Design a new methodology to compare different algorithms according to 34 how often the user watches the TV (length of user profile) if the user prefers blockbuster movies (user preference versus popular or unpopular movies and programs) Design a multi-model system
Metrics 35 Error metrics Mean Square Error (MSE) Root Mean Square Error (RMSE) Mean Absolute Error (MAE) Accuracy metrics Recall Precision Fallout F-measure Both implicit and explicit datasets Only for explicit datasets Top-N recommender systems
Accuracy metrics 36
Accuracy metrics 37
Accuracy metrics 38
Netflix dataset: test user profile 39
Netflix dataset: Global effects algorithm 40 RMSE: 0.95 Recall: 1% F-measure: 0.01
Netflix dataset: Adjusted cosine algorithm 41 RMSE: 1.6 Recall: 8% F-measure: 0.16
Netflix dataset: SVD algorithm 42 RMSE: 2.7 Recall: 17% F-measure: 0.28
Quality evaluation 43 Focus on future performance on new data Proper partitioning of original data set into: training set test set Test set must be different and independent from training set Active user: should be left out of the model 43
Hold-out 44
Leave-one-out 45
K-fold 46
Agenda 47 IPTV architecture Recommender algorithms Evaluation of different algorithms Multi-model systems
Recommender system architecture 48 Inputs Items Content (ICM) Business Rules Real time calls Web Services STB client Batch Processing Real-time Recommendation STB server Users Ratings (URM) Model Repository STB client
Proposed approach Batch system Statistical analysis of the dataset Definition of a number of models Accuracy evaluation for different user profiles Run-time system User profile analysis Selection of best candidate model Recommendation
Multi-model recommender engine
Dataset statistical analysis (example) 51
Dataset statistical analysis (example) 52 Percentage of rated items in the top-rated 1 0.75 0.5 0.25 NM ML NF 0 10 0 10 1 10 2 10 3 10 4 Position of the items in the top-rated
Dataset statistical analysis (example) 53 User groups Item popularity 20 or more Popular 10...19 2...9 Non-Popular
Popular vs. unpopular: SVD algorithm - NF 54 0.35 0.3 0.25 all popular unpopular Recall 0.2 0.15 0.1 0.05 0 0 200 400 600 800 1000 Latent size
Popular vs. unpopular: SVD algorithm - NM 55 0.25 all popular unpopular 0.2 Recall 0.15 0.1 0.05 515 50 100 200 300 Latent size
User profile length NM recall 56 All Popular Unpopular Group SVD Cos -like NBN_S NBN_I NBN_U 2-9 11,21% 19,35% 19,65% 17,23% 21,65% 10-19 11,23% 13,11% 12,09% 13,62% 12,60% 20 -inf 9,91% 8,01% 6,45% 6,65% 6,52% Group SVD Cos -like NBN_S NBN_I NBN_U 2-9 22,21% 31,21% 31,54% 26,17% 33,93% 10-19 24,12% 27,12% 24,61% 27,36% 25,59% 20 -inf 25,72% 22,71% 20,71% 21,14% 20,93% Group SVD Cos -like NBN_S NBN_I NBN_U 2-9 9,92% 0,81% 0,13% 2,64% 1,48% 10-19 10,01% 1,23% 0,19% 0,56% 0,25% 20 -inf 10,14% 0,70% 0,01% 0,10% 0,01% Best average algorithm (item-based) Multi-model (overall) 15,94% 20,92%