IPTV Recommender Systems

1
IPTV Recommender Systems

• Paolo Cremonesi

2
Agenda

• IPTV architecture
• Recommender algorithms
• Evaluation of different algorithms

3
IPTV architecture
Customers
Service Provider
Network Provider
Content Provider
Head end
Set-top-box (decoder)
VOD
4
IPTV architecture

• 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.

5
IPTV Platform Now
CUSTOMERS FACE DIFFICULTIES FINDING THE RIGHT
CONTENT
HUNDREDS LIVE CHANNELS
THOUSANDS VOD ITEMS
CUSTOMER PURCHASES
CUSTOMER FRUSTRATION
6
IPTV Platform with a recommender systems
From this.
Today recommendations, based on your personal
taste, are
To this.
7
Recommender System how it works
USER DATA
USERS TASTE FRUTIONS AND RATINGS
CONTENT METADATA
RECOMMENDER SYSTEM
CONTENT RECOMMENDATIONS
8
Benefits for the Provider

• Increased revenues
• Increase customer spontaneous purchases
• Attract new customers
• 35 of product sales result from recommendation
  system(source Amazon.com)
• Improved customer satisfaction and retention
• Personalize IPTV experience
• Strengthen customer ties
• Optimized business with insights into customer
  tastes
• Marketing and targeted advertising
• Planning of media content catalogs

9
Benefits for the User

• Improved overall usability
• Reduced time to find the right content in VOD
  catalogs
• No more user frustration
• Personalized experience
• Recommendations tailored to user preferences
• Suggests unheard of content fitting customer
  taste
• Enhanced collaborative platform
• Create sense of community
• Involve users by collecting their feedback
  (ratings)
• Reduce the Channel Zapping problem

10
Watch it VOD Recommendation
Recommend media contents based on customers
profile
Average Customers Rating
11
Watch it Channel Recommendation
Recommend live channels based on customers
profile
12
Watch it TV Program Recommendation
Watch the live event
Recommend TV programs based on customers profile
Schedule the recording of the recommended TV
program
13
Watch it Media Content Rating
Collect customers explicit rating right after
content fruition
14
Agenda

• IPTV architecture
• Recommender algorithms
• Evaluation of different algorithms

15
Recommender systems overview

• Problem
• Users face with large amount of data, getting
  confused in the retrieval process
• Objectives
• Recommend users with just a list of relevant
  items

information needs dynamism

• Good items
• All good items

Information Retrieval
Query - based
?
Profile - based
Information Filtering
information sources dynamims
16
Problem formulation
Recommender
Users ratings
Items metadata
Ranked list

• Item1
• Item2
• Item3
• .
• .
• .
• ItemX

Top N
17
Problem solutions
Recommender
Users ratings
Items metadata
CollaborativeFiltering
Content-based Filtering
18
Recommendation techniques
Recommendation algorithm
Similar Items
Collaborative Filtering
Content-based Filtering
Users with similar taste
User based
Item based
19
Collaborative Filtering
User-based similar users rate an item similarly
Item-based similar items are rated by a user
similarly
User-based similar users rate an item similarly
5
4
?
3
2
2
User
Item
Neighborhood
NB similarity means correlation
20
Collaborative Filtering techniques

• For each user, compute a neighborhood by mean of
• Cosine between user vectors (in the items space)
• Pearson Similarity Coefficient
• ..
• Then, for each unrated item, compute its estimate
  rating based on the rate given by users in the
  neighbourhood.
• Alternative and prominent techniques include SVD
  user-rating matrix decomposition, bayesian
  networks,

21
Singular Value Decomposition
diagonal matrix
A
U
S
VT
U
S
VT
A

m x n
A
VkT
Ak
Sk
Uk

m x n
Svd complexity O(min(nm2, mn2))
22
Collaborative Filtering SVD
svd
A
Ak

• Users in rows
• Items in columns

Vk sqrt (Sk)
Uk sqrt (Sk)
pseudo users
pseudo items
cosine
Ak
23
Folding-in

• 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 u Vk Sk-1

Ak
Uk
Sk
Vk
u
u
24
Collaborative Filtering pro cons

• 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)

25
Content-based Filtering
Bag of Words (BOW) representation

• 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)

Word
Item
26
Content-based Filtering pro cons

• 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 an structured content
• Low efficiency of BOW (Bag of words)
  representation
• Very high-dimensional
• Users tastes must be represented as a function of
  the content features to be learnt
• Unable to exploit quality judgments of other users

27
Content-based Filtering techniques
User-item similarity

• Typically after a Latent Semantic Analysis
• Reduces space dimensionality
• Enhances BOW representation

28
Content-based Filtering Latent Semantic Analysis
U
S
VT
U
S
VT
Word
Item

m x n
Latent dimension
Item
29
Content-based Filtering Latent Semantic Analysis
svd
A
Ak

m x n
Vk sqrt (Sk)

• Terms in rows
• Items in columns

Uk sqrt (Sk)
pseudo terms
pseudo items
cosine
Ak
30
Proposed model
1 0 0 0 0 1 0 1 0 0 1 0 1 0 0 1 0 0 0 0 1 0 0 1 0
0 1 1 0 1 0 0 0 0 1 0 1 0 1 0 0 1 0 1 0 1 0 0 0
0 1 0 1 0 0 1 0 0 0 0 0 1 0 1 0 1 0 0 1 0

• Not all zeros mean user does not like an item
• We can change some zeros with a value from CbF

User ratings
Features extractions
CbF
Meta-data
SVD
Content-based Recommendation
CbfSimilarity
Ratings
31
Proposed model
1 0.5 0 0.8 0 1 0 1 0 0 1 0 1 0 0 1 0 0.6 0 0 1 0
0 1 0 0.4 1 1 0 1 0 0 0 0.8 1 0 1 0 1 0.8 0 1 0
1 0 1 0.5 0 0 0 1 0.5 1 0 0 1 0 0 0.6 0 0 1 0 1
0.4 1 0 0 1 0
User ratings
Features extractions
CbF
Meta-data
SVD
Content-based Recommendation
CbfSimilarity
Ratings
CF
SVD
Mixed Recommendation
(Sorted list)Recommendations
32
Proposed model

• The result is a modular and flexible model
• CF may use any algorithm
• Cbf may use any algorithm as well

Features extractions
CbF
Meta-data
SVD
Content-based Recommendation
CbfSimilarity
Ratings
CF
SVD
Mixed Recommendation
(Sorted list)Recommendations
33
Agenda

• IPTV architecture
• Recommender algorithms
• Evaluation of different algorithms

34
Recommender architecture
Resources management
Features extraction
Featuresrepresentation
Items
Storage
Filter
Compute user-item correlation
Items retrieval
Items recommendation
Users management
Infer and learn profile
Interests/tastes representation
Users
35
System Architecture
Offline inputs
Real time calls
Content Data/Metadata
Web Services
IPTV Interface Layer H/A
(A)
(B)
Users Data
Plain Http
Offline 24/7 Processing engine
Real-time Recommendation Engine
Users behaviour Fruitions Ratings(?)
EJB
RecommenderRepository
The recommendation process is phased in two
steps (A) Step one is an offline process used
to analyze the rating data and generate a model
runs in background and updates the model on a
regular basis (B) Step two is a real-time online
process and uses the model built in step one to
respond in real-time to recommendation/personaliza
tion queries.
36
Datasets
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

37
Dataset users

• 2.36 ratings/user

of users
of ratings
38
Dataset movies

• 100 ratings/movie

of movies
of ratings
39
System evaluation (1)

• Consider the user-rating matrix
• Take some positive ratings off (the items took
  off are the sample items and the relative users
  the sample users)
• Run the algorithm and get a sorted list for the
  sample users

sample
1 0 0 0 0 1 0 1 0 0 1 0 1 0 0 1 0 0 0 0 1 0 0 1 0
0 1 1 0 1 0 0 0 0 1 0 1 0 1 0 0 1 0 1 0 1 0 0 0
0 1 0 1 0 0 1 0 0 0 0 0 1 0 1 0 1 0 0 1 0
1 0 0 0 0 1 0 1 0 0 1 0 0 0 0 1 0 0 0 0 1 0 0 1 0
0 1 1 0 1 0 0 0 0 1 0 1 0 1 0 0 1 0 1 0 1 0 0 0
0 1 0 0 0 0 1 0 0 0 0 0 1 0 1 0 1 0 0 1 0
sample
users
users
items
items
40
System evaluation (2)

• For each sample user i, the local metric is the
  position pi of his sample item inside his own
  sorted list
• The more pi is close to 1, the better the
  algorithm is performing for such user i.
• For a dataset, the global metric
• Fix a threshold T (top T rated)
• Count the fraction of sample users with pi lt T

Position pi
user i

• itemC
• itemD
• itemA
• itemB
• itemH

1 0 0 0 0 1 0 1 0 0 1 0 1 0 0 1 0 0 0 0 1 0 0 1 0
0 1 1 0 1 0 0 0 0 1 0 1 0 1 0 0 1 0 1 0 1 0 0 0
0 1 0 1 0 0 1 0 0 0 0 0 1 0 1 0 1 0 0 1 0
1 0 0 0 0 1 0 1 0 0 1 0 0 0 0 1 0 0 0 0 1 0 0 1 0
0 1 1 0 1 0 0 0 0 1 0 1 0 1 0 0 1 0 1 0 1 0 0 0
0 1 0 0 0 0 1 0 0 0 0 0 1 0 1 0 1 0 0 1 0
users
users
items
items
41
System evaluation (3)

• N sample users, each one with one sample item
• Fix a threshold T
• Define
• (pi piltT) ? N p(l)?1
• (pi piltT) ? 0 p(l)?0
• Leave-one-out
• repeat for all sample users and all sample items
• N non zero elements in the rating matrix

42
User-based collaborative Latent size k
43
User-based collaborative number of ratings
Real time parameter
44
User-based collaborative number of ratings
45
Hybrid Content Based Threshold
46
Hybrid Content Based Threshold
47
Evaluation
res1A
48
Evaluation
res1B
49
Evaluation
Res5/6
50
Evaluation
res3A
51
Evaluation
res3B
52
Any questions
?