a rec om mender system for an iptv

8/3/2019 A Rec Om Mender System for an IPTV

1/16

WHITEPAPER March 2010

A Recommender System for an IPTVService Provider: a Real Large-ScaleProduction EnvironmentRiccardo Bambini, Paolo Cremonesi, and Roberto Turrin

www.contentwise.tv


2/16

2

www.contentwise.tv

A RecommendeR SyStem foR An IPtV SeRVIce PRoVIdeR:A ReAl lARge-ScAle PRoductIon enVIRonment

Riccardo Bambini, Paolo Cremonesi, and Roberto Turrin

AbstractIn this chapter

1

we describe the integration o a recommender system into theproduction environment o Fastweb, one o the largest European IP Television (IPTV) providers.The recommender system implements both collaborative and content-based techniques,suitable tailored to the specic requirements o an IPTV architecture, such as the limitedscreen denition, the reduced navigation capabilities, and the strict time constraints.The algorithms are extensively analyzed by means o o-line and on-line tests, showing theeectiveness o the recommender systems: up to 30% o the recommendations are ollowedby a purchase, with an estimated lit actor (increase in sales) o 15%.

1 IntRoductIon

IP Television (IPTV) broadcasts multimedia content (e.g., movies, news programs,documentaries)in digital ormat via broadband Internet networks [11, 7]. IPTV services include scheduledtelevision programs and video on demand (VoD) contents [12]. In the rest o the chapter wewill generically reer to both scheduled television programs and video-on-demand contentsas items.In this chapter we present the integration o the Neptunys ContentWise recommendersystem in Fastweb, the rst company in the world to have launched ully IP-based broadbandTV services, in 2001. Fastweb serves hundreds o thousands o IPTV customers, with a catalogo thousands o multimedia contents. Since 2007 Fastweb is part o the Swisscom group.ContentWise recommender algorithms have been developed with the cooperation o theComputer Science Department at the Politecnico di Milano.

Riccardo Bambinifaswb, via Francesco Caracciolo 51, Milano, Italy

Paolo CremonesiPii i mia, p.zza Leonardo da Vinci 32, Milano, Italy

Roberto Turrinnp, via Durando 10, Milano, Italy

1 This is a drat chapter rom the book Rr Sss Habk editetd by P.B.Kantor, F.Ricci, L.Rokach and

B.Shapira, and published by Springer (ISBN: 978-0-387-85819-7).The book is available rom Springer web site http://www.springer.com/computer/ai/book/978-0-387-85819-7or rom Amazon http://www.amazon.com/Recommender-Systems-Handbook-Paul-Kantor/dp/0387858199


3/16

3

A Recommender System or an IPTV Service Provider

Dierently rom conventional television, IPTV allows an interactive navigation o the availablecontent [5] and, in particular, IPTV allows to collect implicit usage data and explicit userpreerences or providing a personalized user navigation. The user interacts with the IPTV

system by means o a special electronic appliance, reerred to as set-top-box(STB). There aresubstantial peculiarities o the STB that limit the user interaction: (i) users control the STB bymeans o a remote control, which is rather limited in the set o actions it allows to perorm, (ii)the user interace is shown on a TV screen typically designed to be looked at rom a distancelarger than that between a PC and the user, and (iii) the system deals with multimedia content,whose navigation is not particularly ast, mainly because o the channel switching time.Dierently rom traditional e-commerce domains (e.g., Amazon, Netfix, iTunes, IMDB, Last.m) where recommender systems have been exploited, IPTV recommender systems need to

satisy particular requirements:

the list of proposed items has to be small because of the limited screen denitionand the reduced navigation capabilities;

the generation of the recommended itemsmust respect very strict time constraints(ew milliseconds) because TVs customers are used to a very responsive system;

the system needs to scale up in a successful manner with both the number ofcustomers and items in the catalog;

part of the catalog is highly dynamic because of live broadcast channels.

The recommender system deployed in Fastweb generates recommendations by means otwo collaborative algorithms (based on item-to-item similarities and dimensionality reductiontechniques) and one content-based algorithm (based on latent semantic analysis). Therecommender system selects the proper algorithm depending on the context. For instance,i the user is reading a movie synopsis, looking or movies with his preerred actors, thealgorithm used is the content-based one. In order to respect the strict real-time requirements,the recommender system and the underlying algorithms ollow a model-based approach andhave been logically divided into two asynchronous stages, the batch stage and the real-timestage.The input data o the whole architecture is composed by: (i) the item-content matrix and (ii)the user-rating matrix. The item-content matrix (ICM) describes the main attributes (metadata)

o each item, such as the title o a movie, the set o actors and its genre(s). The user-ratingmatrix (URM) collects the ratings (i.e., preerences) o users about items. Ratings are mainlyimplicit, e.g., the system can detect i a user watched a program, without knowing explicitlythe users opinion about that program.Beore deploying the recommender system in production, extensive perormance analysishas been perormed by means ok-old cross validation. The results suggests a 2.5% recallor the content-based algorithm, while the collaborative algorithms are able to reach a recallo more than 20%.The recommender system has been released to production environment in October 2008 andis now available or one o the Fastweb VOD catalogs. The system is actually providing, onaverage, 30000 recommendations per day, with peaks o almost 120 recommendations perminute during peak hours. On-line analysis shows that almost 20% o the recommendationsare ollowed by a purchase rom the users.


4/16

4

www.contentwise.tv

The estimated lit actor (i.e., increase in VOD sales) is 15%.The rest o the chapter is organized as ollows. Section 2 shows the typical architecture oan IPTV provider. Section 3 presents the architecture o the recommender system. Section 4explains the recommender services implemented into the Fastweb IPTV architecture. Section5 evaluates the quality o recommendations. Finally, Section 6 draws the conclusions.

2 IPtV ARcHItectuRe

IPTV, also called Internet Protocol Television, is a video service that delivers high qualitytraditional TV channels and on-demand video and audio contents over a private IP-based

broadband network. From the end users perspective, IPTV looks and operates just like astandard TV service. The providers involved in deploying IPTV services range rom cable andsatellite TV carriers to large telephone companies and private network operators. IPTV has anumber o unique eatures [5]:

Sppr r iraiv tV: dierently rom conventional TV, where the communication isunidirectional, the two-way capabilities o IPTV systems allow the user to interact with thesystem.ti shii: IPTV permits the temporal navigation o programs (e.g., ast orward, pauseand rewind) thanks to the Personal Video Recording (PVR), a mechanism or recording and

storing IPTV content or later viewing.Prsaizai: IPTV allows end users to personalize their TV viewing experience by lettingthem decide what they want to watch and when they want to watch it.

Figure 1 shows a generic IPTV system architecture that supports live broadcast TV channels(also called linear channels) and video on-demand (VOD). Broadcastm TV service consists inthe simultaneous reception by the users o traditional TV channels either ree-to-air or pay-per-view. Video on-demand service consists in viewing multimedia content made available bythe service provider, upon request.

Fig. 1 Architecture o an IPTV system.


5/16

5


The IPTV data center (also known as the headend) receives linear channels rom a variety osources including terrestrial, satellite and cable carriers. Once received, a number o dierenthardware components, ranging romencoders and video servers, are used to prepare thevideo content or delivery over an IP based network. Ondemand contents are stored in aststorage boxes (e.g., using solid-state disks).The set-top-box (STB) is an electronic appliance that connects to both the network and thehome television: it is responsible or processing the incoming packet stream and displayingthe resulting video on the television. The user interacts with the STB by means o a hand-heldremote control. The remote control gives the user access to additional eatures o the STB,such as the Electronic Program Guide (EPG), a listing o available channels and program oran extended time period (typically 36 hours or more).

2.1 IPtV search problems

To benet rom the rich set o IPTV channels and contents, users need to be able to rapidlyand easily nd what they are actually interested in, and do so eortlessly while relaxing onthe couch in their living room, a location where they typically do not have easy access to thekeyboard, mouse, and close-up screen display typical o desktop web browsing. However,searching or a live channel or a VOD content is a challenging problem or IPTV users [4].When watching live television, users browse through a set o available channels until they

nd something interesting. Channel selection (zapping) involves two steps: (a) sampling thecontent to decide whether to continue or stop watching the channel, and (b) switching acrossmultiple channels or repeated sampling, until a desired channel is ound. The problem oquickly nding the right channel becomes harder as the number o channel oerings grows in

modern IPTV systems. Moreover, IPTV channel switching time is not particularly responsive,compared to traditional TV, because o technological limitations [9].When searching or VOD content, IPTV users generally have to either navigate an complex,pre-dened, and oten deeply embedded menu structure or type in titles or other keyphrases using an on-screen keyboard or triple tap input on the remote control keypad. Theseinteraces are cumbersome and do not scale well as the range o content available increases.Moreover, the television screens usually oer a limited resolution with respect to traditional

personal computer screens, making traditional graphical user interaces dicult to use.This diers rom traditional web-based domains (e.g., e-commerce web sites), where thecontent is textual, suited or inormation categorization and keywordbased seek and retrieval,and the input devices (keyboard and mouse) allow to point an arbitrary object on the screenand to easily enter text.The integration o a recommender system into the IPTV inrastructure improves the userexperience by providing a new and more eective way o browsing or interesting programsand movies. However, such integration has to deal with the ollowing issues:

User identifcation. The STB is used indistinctly by all the components o a amily, and theIPTV recommender system can not identiy who is actually watching a certain program.Ra-i rqirs. The IPTV recommender systems must generate recommendationswithin very strict real-time constraints (ew milliseconds) in order to avoid a slow down o


6/16

6

www.contentwise.tv

the user navigation, already aected by the long channel switching time.Qai aaa. Dierently rom web-based domains, content-based IPTVrecommender algorithms makes use o low-quality metadata. This aspect is particularlyevident with live channels, where new content is added every day at a very high rate, andthe only available metadata that can be used to describe programs can be ound in the EPG(electronic program guide).ecommender algorithms makes use o low-quality metadata.This aspect is particularly evident with live channels, where new content is added every dayat a very high rate, and the only available metadata that can be used to describe programscan be ound in the EPG (electronic program guide).

3 RecommendeR SyStem ARcHItectuRe

The architecture o the Fastweb recommender system is shown in Figure 2.

3.1 Data collection

The logical component in charge o pre-processing the data and generating the input othe recommender algorithm is reerred to as data collector. The data collector gathers datarom dierent sources, such as the EPG or inormation about the live programs, the content

provider or inormation about the VOD catalog and the service provider or inormationabout the users.

The Fastweb recommender system does not rely on personal inormation rom the users(e.g., age, gender, occupation). Recommendations are based on the past users behavior(what they watched) and on any explicit preerence they have expressed (e.g., preerred

genres). I the users did not speciy any explicit preerences, the system is able to iner themby analyzing the users past activities.An important question has been raised in Section 2: users interact with the IPTV systemby means o the STB, but typically we can not identiy who is actually in ront o the TV.Consequently, the STB collects the behavior and the preerences o a set o users (e.g.,

Fig. 2 Architecture o the recommender system ContentWise.

EPG

Contentprovider

Datasource

Data collectorSTB

server

STB client

STB client

.......Recommender

algorithms

Recommenderweb services

Serviceprovider

Repository

ICM URM


7/16

7


the component o a amily). This represents a considerable problem since we are limited togenerate per-STB recommendations. In order to simpliy the notation, in the rest o the paperwe will reer to user and STB to identiy the same entity. The user-disambiguation problemhas been partially solved by separating the collected inormation according to the time slotthey reer to. For instance, we can roughly assume the ollowing pattern: housewives use towatch TV during the morning, children during the aternoon, the whole amily at evening,while only adults watch TV during the night. By means o this simple time slot distinction weare able to distinguish among dierent potential users o the same STB.Formally, the available inormation has been structured into two main matrices, practicallystored into a relational database: the item-content matrix (ICM) and the user-rating matrix(URM).

The ormer describes the principal characteristics (metadata) o each item. In the ollowingwe will reer to the item-content matrix as W, whose elements wci represent the relevanceo characteristic (metadata) cor item i. The ICM is generated rom the analysis o the seto inormation given by the content provider (i.e., the EPG). Such inormation concerns, orinstance, the title o a movie, the actors, the director(s), the genre(s) and the plot. Note thatin a real environment we can ace with inaccurate inormation especially because o the ratenew content is added every day. The inormation provided by the ICM is used to generate acontent-based recommendation, ater being ltered by means o techniques or PoS (Part-o-Speech) tagging, stop words removal, and latent semantic analysis [13].Moreover, the ICM can be used to perorm some kind o processing on the items (e.g.,

parental control).The URM represents the ratings (i.e., preerences) o users about items. In the ollowing wewill reer to such matrix as R, whose elements rpi represent the rating o user p about item i.Such preerences constitute the basic inormation or any collaborative algorithm. The user

rating can be either explicit or implicit, according to the act that the ratings are explicitlyexpressed by users or are implicitly collected by the system, respectively.Explicit ratings condently represent the user opinion, even though they can be aected bybiases [1] due to: user subjectivity, item popularity or global rating tendency. The rst biasdepends on arbitrary interpretations o the rating scale. For instance, in a rating scale between1 and 5, some user could use the value 3 to indicate an interesting item, someone else coulduse 3 or a not much interesting item. Similarly, popular items tend to be overrated, while

unpopular items are usually underrated. Finally, explicit ratings can be aected by globalattitudes (e.g., users are more willing to rate movies they like).On the other hand, implicit ratings are inerred by the system on the basis o the user-systeminteraction, which might not match the user opinion. For instance, the system is able tomonitor whether a user has watched a live program on a certain channel or whether theuser has uninterruptedly watched a movie. Despite explicit ratings are more reliable thanimplicit ratings in representing the actual user interest towards an item, their collection canbe annoying rom the users perspective.

The current deployment o the Fastweb recommender system collects only implicit ratings,but the system is thought to work when implicit and explicit ratings coexist. The rating scaleis between 1 and 5, where values less than 3 express negative ratings, values greater or equalto 3 express positive ratings. In absence o explicit inormation, the rating implicitly inerredby monitoring the user behavior is assumed to be positive (i.e., greater than 3).


8/16

8

www.contentwise.tv

In act, whether a user starts watching a certain program, there must be some characteristic othis program appealing or the user (e.g., actors or genre). The act that in well-know, explicitdatasets, such as Netfix and Movielens, the average rating is higher than 3, motivates thisassumption. The system treats dierently live IPTV programs and VOD content:

IPtV prras. The rating is proportional to the percentage user play time (e.g., [8, 15]), i.e.,the percentage o program the user has watched. Let us assume L is the program play timeand tis the user play time. Play times less than 5 minutes are discarded. I a user watchedthe entire program the rating is 5, i the user watched 5 minutes the rating is 3, otherwise therating is a value between 3 and 5 given by:

(1)

where t and L are expressed in minutes.At this early stage o the project, the main goal is not to dene an accurate implicit ratingmechanism, but rather, to lter out noisy inormation (e.g., TV channel zapping).VOD movies. When watching a VOD movie, users explicitly request to buy and to pay orthat movie. For that reason, independently rom the user play time, when a user requests a

VOD movie, the system assign an implicit ratings equals to 4.

As aorementioned, should Fastweb start collecting explicit ratings too, they will naturallycoexist with implicit ratings in the URM.The ratings stored in the URM, beore being used by the recommender algorithms, arenormalized by subtracting the constant value 2.5. This allows the algorithms to distinguishbetween positive and negative ratings, because values greater or equals to 2.5 (i.e., 3, 4, and5) remain positive, while values less than 2.5 (i.e., 1 and 2) become negative. In the rest othe chapter we will assume that the recommender algorithms receive as input a normalizedURM.Finally, users can express explicit preferences about the content they would like to watch. Forinstance, by means o the graphical interace, a user can set his preerred actors. The content-based algorithmtakes into consideration such inormation and biases the recommendedmovies toward the expressed preerences.

3.2 Batch and real-time stages

The recommender algorithms process the ICM and the URM described in Section 3.1 andthey interace with the STB server by means o web services, as shown in Figure 3.In order to respect the strict real-time requirements, the recommender system and theunderlying algorithms ollow a model-based approach [14, 3], i.e., they rst develop amodel o the user ratings and/or o the items, then they compute the recommendations.

Consequently, the algorithms have been logically divided into two stages, the batch stageand the real-time stage:


9/16

9


the batch stage creates a low dimensional representation (i.e., a model) of the input data.It is usually executed during the service o-peak hours, with a requency which dependson the rate new items/users are added into the system (e.g., once a day);

the real-time part uses the model in order to serve calls coming from the web servicesinterace and satisying the real-time constraints. The system output can be urther constrainedby post-processing, marketing rules (e.g., pushing up some movies, or ltering some channels).

The model repository makes the two stages asynchronous, e.g., while the real-time stage is

recommending users by using a certain model, the batch stage can compute a new, updatedmodel.Despite such logical division o the recommending process, the model construction in real

domains can still be challenging because o input data size and the related time and memoryrequirements. For this reason, we have implemented highperorming, parallel versions o themost demanding matrix operations, optimized or sparse and big matrices, such as: matrix-matrix and matrix-vector multiplication, matrix transposition, column/row normalization,and singular value decomposition (svd). In particular, svd has been used with two o thethree recommender algorithms (one content-based and one collaborative), allowing togreatly reduce the space dimensionality, with benets both in terms o memory and timecomplexity.

As we will show in the ollowing sections, by its own, svd denes a model o the data, cleaningup the data noise and strengthening the correlation among similar inormation.Realistic datasets with millions o users and items can have in principle prohibitive memoryrequirements. Fortunately, matrices such as URM and ICM are typically very sparse. In act,most o the users interact (e.g., rate or watch) with very ew items compared with the size

o the catalog (e.g., the average users have watched ew dozens o movies in a catalog othousands). Sparse matrices can be treated using very ecient representations. Note that,even though such matrices aren sparse, we could have diculties in maintaining the data inmemory. For such reasons, we opted or a solution based on a sort omemory virtualization,similar to the swap capability o any operating systems. Dierently rom the operating systemvirtual memory, our virtualization policy is tailored ad-hoc or each matrix operation, in orderto limit the data exchange between memory and storage.

Items Content(ICM)

Inputs

Batchprocessing

Batchprocessing

STB

server

BusinessRules

Recommenderweb services

Users Rating(URM)

ModelRepository

Fig. 3 Recommender system: batch and real-time stage.


10/16

10

www.contentwise.tv

4 RecommendeR SeRVIceS

This section presents the implemented recommender services and how they impact into theuser interaces and the IPTV architecture. The recommender system can generate both content-based and collaborative-based recommendations. As summarized in Figure 4, content-basedalgorithms are applied both to VOD and live TV domains, while collaborative algorithms areapplied only to VOD.At the current stage o the integration, Fastweb is exposing the ull seto recommender services to a selected set o beta test users beore the eective release.

The other customers have access to a reduced set o the recommender unctionalities. An

example o the user interace available by means o the STB is shown in Figure 5.The services released to the ull customer base concern one o the catalog o VOD domain.Recommendations are provided by a content-based algorithm based on Latent SemanticAnalysis. The content-based algorithm has been preerred or the rst ew months o activitybecause collaborative algorithms suer rom the cold-start problem. Moreover, collaborative

recommenders need to record the behavior o users. This aces Fastweb with delicate legalquestions that require, or instance, to obtain authorizations rom customers or storing andmanaging their data, and to implement solutions to grant condentiality and anonymity osuch inormation.

Fig. 4 Application o recommender algorithms to VOD and live TV.

Fig. 5 Recommender system user interace

Item-based CF

SVD-CF

LSA-CB

VOD Live TV


11/16

11


5 SyStem eVAluAtIon

In this section we rst discuss the quality o the recommender system by means o accuracymetrics computed adopting an o-line testing. We later present some eedbacks rom theon-line analysis o the recommender system.The o-line tests are based on the views collected during 7 months o users activity romone o the VOD catalogs. Figure 6 shows the evolution o the number o views. The averagenumber o views per days is about 1600, with up to 3300 views during week-end days. Figure7, 8, and 9 complete the analysis by showing the time evolution o, respectively, the numbero active users, the number o active items, and the dataset density. Active users are usersthat have rated at least one item. Similarly, active items are items that have been rated by

at least one user. The dataset density is the ratio between the number o ratings and theproduct o the number o active users and the number o active items. We can notice romFigure 9 that the trend is not monotone. In act, when a new user watches her/his rst item,we have one new active user, and the dataset decrease its density.

5.1 Off-line analysis

Typical approaches or recommender system evaluation are based either on error metrics(e.g., RMSE and MAE) [10] or classication accuracy metrics (e.g., recall, precision, and all-

out) [6, 2]. Since we only dispose o implicit ratings, expressing positive user interests, we arepractically constrained in evaluating the quality o the system by means o accuracy metrics.To this end, Table 1 presents the recall o the three ContentWise algorithms2, respectively:LSA-CB, item-based-CF and SVD-CF.Recall is oten used in inormation retrieval, where it species the percentage o relevant itemsthat have been retrieved by, or instance, a search engine. In our domain, recall indicates howmany movies that users have eectively watched are recommended by the recommenderalgorithm. To this purpose, we ollow a leaveone- out approach:

for each user in the test set, we select one rated item

2 To obtained a detailed report, please XXXX

Fig. 6 Number o views collected during 7 months o users activity rom one o the VOD catalogs.

0 20 40 60 80 100 120 140 160 180 2000

0.5

1

1.5

2

2.5

3

3.5x 10

5

Elapsed days

Numberofviews


12/16

12

www.contentwise.tv

Fig. 7 Number o active users rom the same VOD catalog.

Fig. 8 Number o active items rom the same VOD catalog.

Fig. 9 Evolution o rating density in the same VOD catalog. Density is computed by considering the number o views (i.e.,ratings) with respect to the number o active users and active items.

0 20 40 60 80 100 120 140 160 180 2003

4

5

6

7

8

9

10x 10

-3

Elapsed days

Density

0 20 40 60 80 100 120 140 160 180 200250

300

350

400

450

550

600

700

Elapsed days

Numberofactiveitems

x 104

0 20 40 60 80 100 120 140 160 180 2000

1

2

3

4

5

6

Elapsed days

Numberofactiveitems


13/16

13


the selected item is removed from the user prole, and we generate a recommendationor this modied user prole; items already rated by the user are ltered out rom therecommendation list.

if the removed item is recommended within the rst 5 positions we have a hit, i.e., amovie that has been watched by a user has been recommended by the algorithm

(accordingly to the Fastweb user interace, the recommended list is limited to 5 items); the process is repeated for each item and for each user.

The recall is the percentage o hits with respect to the number o tests.The test set is selected dierently according to the kind o algorithm. In act, content-basedalgorithms build their model by using the ICM, and the test set can be the whole URM. On

the other hand, the model o collaborative algorithms is based on the URMitsel, so theyhave been evaluated with a 10-old cross validation approach, i.e., we have randomly splitthe users into 10 olds and, in turn, one old has been used as test set or computing therecall, while the remaining nine olds have been used to generate the model. Each test oldis analyzed by means o the leave-one-out approach. The reported results are the averagerecall among the 10 olds.The table reports the recall o the recommender algorithms both ater 3 months o activityand ater 6 months o activity, showing the time evolution o the system. Furthermore, thequality o recommendation o the three algorithms are compared with a trivial algorithm, usedonly or comparison purposes: the top-rated. The toprated algorithm is a basic collaborative

algorithm that recommends to any user a x list o items, ordered rom the most popular tothe less popular (discarding items already rated by the user).Table 1 shows that, in some cases, the quality o recommendations ater 6 months is lowerthan ater 3 months. We would expect that, as the system collects ratings, it acquires more

precise user proles and the quality o recommendations should improves. However, ater3 months o activity there are 510 active items, while ater 6 months we have 621 activeitems. In terms o probability, ater 3 months an algorithm has to pick up 1 items among 510candidates, while ater 6 months the number o candidates is 621, as shown in Figure 8.

0 20 40 60 80 100 120 140 160 180 2001

2

3

4

5

6

7

Elapsed days

Averageprofilelength

Fig. 10 Time evolution o the average user prole length. Prole lengths are computed on users active in one o the VODcatalogs.


14/16

14

www.contentwise.tv

5.2 On-line analysis

In this section we integrate the previous results, obtained rom an o-line analysis o therecommender algorithms, with an on-line analysis, i.e., we directly study the eedback onthe running recommender system. As explained in Section 4, the reported data reer to thecontent-based algorithm applied on one o the VOD catalogs.In order to empirically evaluate the recall, we assume that whether a user watches a movieater it has been recommended by the system, such movie is relevant or the user and thisrepresents a success or the recommender system.Let us dene the recommendation success, which measure the number o movies that havebeen viewed within a certain time period ater being recommended. Indicating with b(t)the

recommendation success and with w(t)the number o movies watched by the same userswithin a time period t rom a recommendation, we can compute an empirical recallas thepercentage ratio between the recommendation success and the number o views:

(2)

The empirical recall represents the percentage o views that have been triggered by therecommender algorithm. The specied indexes depend on the time period t that is takeninto consideration ater the recommendation has been provided to the user. Please note that

a too long time period t could loose the dependency between the recommendation and theview. Table 2 shows the average quality o the system computed by monitoring the viewswithin 2 hours, within 24 hours, and within 7 days rom the recommendation. The reportedresults distinguish between popular and unpopular items.From the table we can observe that the empiric recall is larger or unpopular movies withrespect to popular movies. In act, popular movies are already known by users, even withoutbeing suggested by the recommender system. For instance, either the user has alreadywatched a popular movie (e.g., at cinema) or it is not interested in it.

Algorithm

Collaborative Item-Based

Collaborative SVD

Content LSA

Top-rated

Recall3 months 6 months

14.0% 12.6%

12.0% 12.0%

2.6% 2.6%

0.4% 1.0%

Table 1 Recommendation quality in one o the VOD catalogs.


15/16

15


6 concluSIonS

The integration o the ContentWise recommender systems into the Fastweb architecture

positively impacts both the customers and the service provider. Three major considerationsderive rom the on-line analysis, conrming the positive eects o the recommender system:(i) users preers to browse the VOD catalog by means o the recommender interace, (ii) userstend to watch recommended movies within ew hours, and (iii) users increase the number owatched movies.Further experiments are currently running on the other catalogs o Fastweb, testing andtuning the quality o all the implemented recommender algorithms and monitoring the cold-start phase o the system in order to complete the release o recommender services. Otherongoing works are addressing the problem o accurately estimating implicit ratings rom userbehavior.

RefeRenceS

1. Bell, R.M., Koren, Y.: Scalable collaborative ltering with jointly derived neighborhood

interpolation weights. 7th IEEE Int. Con. on Data Mining pp. 4352 (2007)2. Cremonesi, P., Lentini, E., Matteucci, M., Turrin, R.: An evaluation methodology or

recommender systems. 4th Int. Con. on Automated Solutions or Cross Media Contentand Multi-channel Distribution pp. 224231 (2008)

3. Deshpande, M., Karypis, G.: Item-based top-n recommendation algorithms. ACMTransactions on Inormation Systems (TOIS) 22(1), 143177 (2004). DOI http://doi.acm.

org/10.1145/ 963770.9637764. Geneve, U.D., Marchand-maillet, S.: Vision content-based video retrieval: An overview5. Hand, S., Varan, D.: Interactive narratives: Exploring the links between empathy,

interactivity and structure pp. 1119 (2008)6. Herlocker, J., Konstan, J., Terveen, L., Riedl, J.: Evaluating collaborative ltering

recommender systems. ACM Transactions on Inormation Systems (TOIS) 22(1), 553(2004)

7. Jensen, J.F.: Interactive television - a brie media history 5066, 110 (2008)

8. Kelly, D., Teevan, J.: Implicit eedback or inerring user preerence: a bibliography.SIGIR Forum 37(2), 1828 (2003). DOI http://doi.acm.org/10.1145/959258.959260

9. Lee, Y., Lee, J., Kim, I., Shin, H.: Reducing iptv channel switching time using h.264scalable video coding. Consumer Electronics, IEEE Transactions on 54(2), 912919(2008). DOI 10.1109/TCE.2008.4560178

2 hours

17.0%

24 hours 7 days

19.8% 24.70%

Table 2 Average empiric recall on the considered VOD catalog. Results reer to three time periods ater the recommendation(2 hours, 24 hours, and 7 days) and are separated between popular and unpopular movies.


16/16

16

www.contentwise.tv

10. Powers, D.M.W.: Evaluation: From Precision, Recall and F-Factor to ROC, Inormedness,

Markedness and Correlation (2000)11. Raey, R.A., Gibbs, S., Hoch, M., Gong, H.L.V., Wang, S.: Enabling custom

enhancements in digital sports broadcasts pp. 101107 (2001). DOI http://doi.acm.

org/10.1145/363361.36338412. Sun, J., Gao, S.: Iptv based on ip network and streaming media service station. MIPPR

2007: Remote Sensing and GIS Data Processing and Applications; and InnovativeMultispectral Technology and Applications 6790(1), 67904Q (2007).DOI 10.1117/12.749611. URL http: //link.aip.org/link/?PSI/6790/67904Q/1

13. Van Rijsbergen, C.J.: Inormation Retrieval, 2nd edition. Dept. o Computer Science,University o Glasgow (1979). URL citeseer.ist.psu.edu/vanrijsbergen79inormation.

html14. Vozalis, E., Margaritis, K.: Analysis o recommender systems algorithms. Proc. o the 6th

Hellenic European Con. on Computer Mathematics and its Applications (2003)15. Zhang, H., Zheng, S., Yuan, J.: A personalized tv guide systemcompliant with mhp.

Consumer Electronics, IEEE Transactions on 51(2), 731737 (2005). DOI 10.1109/TCE.2005.1468026

a rec om mender system for an iptv

Documents