TV Predictor: Personalized Program Recommendations To
TV Predictor: Personalized Program Recommendations tobe displayed on SmartTVsChristopher KraussLars GeorgeStefan ArbanowskiFraunhofer Institute FOKUSKaiserin-Augusta-Allee 31Berlin, GermanyHumboldt-Universität zu BerlinUnter den Linden 6Berlin, GermanyFraunhofer Institute FOKUSKaiserin-Augusta-Allee 31Berlin, fer.deABSTRACT1.Switching through the variety of available TV channels tofind the most acceptable program at the current time canbe very time-consuming. Especially at the prime time whenthere are lots of different channels offering quality content itis hard to find the best fitting channel.This paper introduces the TV Predictor, a new application that allows for obtaining personalized program recommendations without leaving the lean back position in frontof the TV. Technically the usage of common Standards andSpecifications, such as HbbTV, OIPF and W3C, leveragethe convergence of broadband and broadcast media. Hintsand details can overlay the broadcasting signal and so theuser gets predictions in appropriate situations, for instancethe most suitable movies playing tonight. Additionally theTV Predictor Autopilot enables the TV set to automatically change the currently viewed channel. A Second ScreenApplication mirrors the TV screen or displays additionalcontent on tablet PCs and Smartphones.Based on the customers viewing behavior and explicitgiven ratings the server side application predicts what theviewer is going to favor. Different data mining approachesare combined in order to calculate the users preferences:Content Based Filtering algorithms for similar items, Collaborative Filtering algorithms for rating predictions, Clustering for increasing the performance, Association Rules foranalyzing item relations and Support Vector Machines forthe identification of behavior patterns. A ten fold cross validation shows an accuracy in prediction of about 80%.TV specialized User Interfaces, user generated feedbackdata and calculated algorithm results, such as AssociationRules, are analyzed to underline the characteristics of sucha TV based application.With the rapid growth of available content via the internet the users are not able to consume all offered products atonce - nor in a life time. So they have to find a consideredselection of media items they want to consume. A Recommendation System helps its users to find a pre selected list ofitems they might be interested in. Therefore these enginesuse a set of different approaches to get a prediction of theusers’ interests.In a closed system like a web platform there exists a fixedamount of items and a group of users is trying to find thebest fitting item within this platform. In general an itemcan be every media a user is searching for. This can be forinstance a text document, a picture, an audio file, an ondemand video file or a live TV program.The term SmartTV (also called hybrid TV) is a madeup word in allusion to the term SmartPhone and has noofficial definition. Today TV sets are called hybrid TVs,when they perform more than just showing moving picturesor teletext. For instance they are able to render websitesoverlaying the regular TV program as they are connected tothe internet. The Hybrid Broadcast Broadband TelevisionStandard (HbbTV) allows for enriching the regular linearTV signal with an Application-URL and so users can openspecific websites by pressing the red button on their remotecontrols. As broadcasters are only able to enrich their ownsignals, this CE-HTML based website is called broadcasterdepended App.The Fraunhofer Institute for Open Communication Systems (FOKUS) has developed a system, called TV Predictor, that uses the benefits of both technologies, SmartTVsand Recommendation Engines, and so allows for overlayingthe linear TV program with personalized recommendationsand rating predictions. It consists of a typical client-serverarchitecture in which the thin client is only used for rending the User Interface and delegating the user inputs to theserver side recommendation engine. The Java-based serveris able to calculate program recommendations for differentsituations depending on the users request.The Personalized Program Guide (PPG) is a feature ofthis recommendation system allowing to create a ”personalchannel”. It is not an actual TV channel, but an aggregation of the best fitting parts of all available media items. AsHbbTV allows for changing to specific channels, the HbbTVfrontend can automatically change the channel for the userwithout further user input. This results in a lean back situation for the user where the TV automatically shows theKeywordsHybrid TV, SmartTV, recommendation, algorithms, contentbased Filtering, collaborative Filtering, offline, online.Permission to make digital or hard copies of all or part of this work forpersonal or classroom use is granted without fee provided that copies are notmade or distributed for profit or commercial advantage and that copies bearthis notice and the full citation on the first page. Copyrights for componentsof this work owned by others than ACM must be honored. Abstracting withcredit is permitted. To copy otherwise, or republish, to post on servers or toredistribute to lists, requires prior specific permission and/or a fee. Requestpermissions from Permissions@acm.org.BigMine’13 , August 11 - 14 2013, Chicago, IL, USACopyright 2013 ACM 978-1-4503-2324-6/13/08 . 15.00.INTRODUCTION
best fitting program at any time.In order to explain the data mining aspects of such a recommendation system, this paper is structured into 6 sections. The second section introduces basics of hybrid TVsets and related recommendation projects. The third sectionwill focus on the use cases, architecture and main components of the TV Predictor. Afterwards the main recommendation aspects as well as the theory becomes determined,followed by an evaluation. A summary and future researchwill conclude this paper.2.STATE OF THE ARTIn this section some basics for a better comprehensionof this paper will be elucidated. Thereby this paper willrefer to standards of the consumer electronics and differentapproaches for recommender systems.2.1Hybrid TVThe hybrid TV is the result of the increasing convergencebetween television and internet. The TV is not necessarilyconnected to the world wide web directly. It can also beupgraded by incorporating so called set top boxes, additionaldevices as middleware.The internet connection of hybrid television is often onlyused to deliver additional contents to the device and not thewhole live program. The live program in contrast (called linear TV) is delivered using the Digital Video Broadcasting(DVB) standard. In general this DVB signal is still delivered by satellite, cable or in a terrestrial way. SmartTVapplications must be accepted by the broadcaster. So theycan guarantee that the web application is optimized for interpretation on a television set and only shows appropriatecontent. Therefore the application URL has to be enteredin the Application Information Table (AIT) of the DVBstream. This is an additional digital container to deliverapplications. This signal containing the TV program andthe application will be received, decrypted and interpretedby the corresponding connected device. So a consumer isable to use the application functions directly on his TV set.In order to fulfill all requirements for displaying and interacting with web pages on TV screens, a specification waspublished in June 2010 by The European Telecommunications Standards Institute (ETSI), called Hybrid BroadcastBroadband Television. The HbbTV specification is basedon different standards of the Open IPTV Forum (OIPF),the Consumer Electronics Association (CEA) and the DVBProject (DVB). Consumer Electronics – HTML (CE-HTML)is a TV web technology based on current W3C standards,like XHTML 1.0, CSS TV Profile 1.0 and JavaScript. In addition this technology allows for using native TV functions,such as changing the TV channel, by using OIPF components.2.2Recommendation SystemsToday there are a lot of Web 2.0 services providing recommendation engines. In Germany only few of the leadingservices are popular or known. First of all the big players likeFacebook, Youtube, Amazon and MySpace are often in use.New emerging SVoD and TVoD services, such as Maxdome,Lovefilm and Watchever, didn’t play a big role in Germanywhen starting this project in November 2011. Today theytry to fill this gap – even on SmartTVs, using HbbTV ornative applications – but with mostly Content Based Rec-ommendation Engines predicting similar videos. So therewas a lack of personalized Recommendation Systems for TVprograms to be displayed on SmartTVs in Germany.Other services focus on the US market, for instance Huluand Netflix. Moreover the US company Rovi demonstrateda white label solution in beta state for TV program guides(cf. [6]) to be shown on TV sets. Academic recommendationservices such as queveo.tv (cf. [2]) or the Content-boostedCollaborative Filtering (cf. [16]) have introduced possiblesolutions to recommend TV related content, but only in aregular web 2.0 environment.It seemed that a Recommendation System predicting TVprograms directly on the target device, that is easy to use,based on the user behaviour and automatic feedback wasneeded in Germany.3.TV PREDICTOR SOLUTIONThe TV Predictor is a German cooperation project between Arvato RTV (content provider and subsidiary company of Bertelsmann) and the Fraunhofer FOKUS. This prototype was introduced at the IFA Consumer Electronics inSeptember 2012 and was designed to recommend programson SmartTVs and connected Second Screen devices. SinceDecember 2012 a productive system was established to beused for free in the regular desktop web environment on thewebsite:http:\\www.rtv.deThe recommendation system uses a set of different criteria to make recommendations which correspond to the users’viewing behavior. When users watch TV Predictor enabledchannels, they can open the recommendation menu by pressing the according button on their remote control. A set ofthe best and most relevant programs for the current user willbe shown. These personalized recommendations are basedon the automatically tracked viewing behavior and explicitlydefined program ratings of the registered user or - in casethey did not sign up - they will get averaged or well-selectedrecommendations.In order to generate the best and most accurate recommendations, the recommendation system combines the bestfitting algorithms in a hybrid way. The usage of these algorithms depends on the user’s request: Find similar programs to the selected one by usingcommon content-based filtering algorithms, such as theCosine Similarity, and by using unsupervised learningalgorithms, such as Association Rules Get program highlights for a specific time period basedon the favorite programs of similar users (Pearson Correlation Coefficient) and predictions of program ratings (Slope One) Calculate a personalized program guide changing thechannel automatically by using Clustering to pre-selectprograms best fitting the user’s interests and ratingpredictions Overlay upcoming program recommendations while watching TV based on recognized behavior patterns (calculated by a Support Vector Machine) to find user interests, such as genres and categories, favored actors,directors and producers or even the preferred channels,weekdays or times to watch specific content
3.1TV specialized User InterfaceThe User Interface on a SmartTV differs from other devices (such as PCs, tablets and Smartphones) in a lot offeatures. The according criteria are focussing on the navigation with a remote control and the display restrictions forolder devices and distant viewers. It is based on CE-HTML,CSS 2 and JavaScript.The TV Predictor Menu will open when users press thered button on their remote control. As shown in figure 1there is a main navigation on the top of the screen. Thepart below is used to display the contents of the currentsection – in this case the program highlights of the currentday.REST-APIs. The interfaces are classified into the 5 maingroups: SubmitClientInput to provide automatic and manually inputs, GetRecommendations to retrieve recommendations, DoCronJobs to activate operations in continuous timeintervals, EngineSettings to adjust the engines behavior andEngineStatistics to offer graphical analysis. Each Interfaceis a Java Servlet. The return type depends on the requesteddata and can be XML or CSV formated data.Figure 1: TV Predictor MenuThe top 7 programs are displayed in the content area. Acircle indicates the users affection for the shown program.This is visualized by a number in the middle of the circle inthe range of [1, 10] representing the prediction value, where1 is the lowest and 10 is the best value. A dark blue circlepoints to an already rated program and a light blue circleshows a real predicted rating.The content menu opens when the users select a specificrecommendation. On the following screen users can get detailed information about the program contents, rate them,see the recommendation value and where applicable get anexplanation of the predicted rating. Moreover they can askfor similar items calculated on demand.The colored keys of the remote control allow to log intothe application (green button), to connect the TV set with asecond screen device (yellow button) and to hide the application and watching the linear program again (red button).Basically the TV Predictor was designed to be shown onSmartTVs, but when users want to watch the linear program again, they can push the contents to a tablet PC or aSmartphone and control the same User Interface by touching on the corresponding screen element. Alternatively theapplication can be mirrored on the TV set and the secondscreen device in order to provide a smart remote controlapplication.3.2Software ArchitectureAccording to figure 2, the client consists of four differentfront end modules: the Consumer Electronics User Interface, the Second Screen Interface, the statistics front endallowing for analyzing user and usage behaviour, as well asthe admin console that allows the administrator to set upand adjust the server system. The server offers 18 differentFigure 2: TV Predictor - ArchitectureThe server infrastructure is hosted in the Amazon cloudusing a load balancer, scalable engine instances and distributed database nodes. The hardware costs are comparatively low as one virtual machine (Amazon EC2 server:c1.medium) can handle up to 400 users at the same time.Another tiny instance (Amazon RDS class: db.m1.small) isneeded for the Amazon Relational Database Service. In order to allow a synchronization of the TV set and the secondscreen device, node.js as a middleware server infrastructurewas introduced.3.3Domain ModelThe core of the recommendation engine focuses on program recommendations. So the core of the domain modelfocuses on it as well. Figure 3 shows an Entity Relationship Model (ERM) containing Users, Programs, their relation and meta data of a Program. A Program on television,such as a series episode, has a lot of meta data. It does notmatter if it is a moving picture, a series episode, news ora documentation, a program can be identified by its content. It may belong to one or more Categories and Genres.Each is encapsulated by an association (ProgramHasCategory and ProgramHasGenre) as it is the rule for ERMs. Ofcourse there are a lot of people involved in the productionof a program, such as the author, the director or the actors.
find similar items. Therefore an element is just comparedto another element by respecting their content informationon meta data. So an item for instance can be very similaror dissimilar to another item. Therefore each attribute hasto be explored and compared to the according attributes ofthe other elements. The Cosine-based Similarity is used tocalculate the similarity of two elements by treating them asvectors: 1, E 2) cSim(e1 , e2 ) cos(E 1 · E 2E 1 E 2 E(1)e1 and e2 are the elements to be compared like items or 1 and E 2 are vectors representing all features ofusers. Ethis element, so when n is the number of all attributes of an 1 (a1,e1 , a2,e1 , a3,e1 , ., an,e1 ). (cf. [4, p.element, then E619], [18, p. 929])Content Based Filtering is used to find the best fittingitem to the current, for instance other videos or even advertisements similar to the watched content. But feedbacks likeuser-item relations are not directly provided by these algorithms. Approaches made on the basis of user-item relationsare called Collaborative Filtering.4.2Figure 3: SmartTV Predictor – ERM of the userprogram relationshipThose people are defined as CrewMembers.Besides a Program may belong to a Series. That is notonly the case if it is a regular TV series, but also if a movieis part of a multi-parted TV event, like a trilogy, or it isrepeated like daily news. A Program can be broadcastedon multiple Channels at multiple times. This constructionis defined as Slot representing a time interval on a specificChannel. So a Program can be shown on different Slots, forinstance a specific episode of a series is broadcasted between10:00 p.m. and 10:30 p.m. on one channel and between 3:00p.m. and 3:30 p.m. on another channel.Primarily a User does not give feedback to a Program, butto a Slot. That is the direct relation to the content whenusers watch TV. What kind of Program a User is interestedin will be analyzed by Filtering algorithms afterwards thuspreparing additional relations between users and situations.4.RECOMMENDATION ALGORITHMSThe TV Predictor Recommendation Engine has the goalto filter the most relevant items from the set of all items.Top-N Filtering means that the resulting collection of elements does not contain the whole data set, but only the firstN elements of an ordered list. So only the Top-7 programs(e.g. with highest predicted rating) are finally offered to auser.4.1Content Based FilteringIn the TV Predictor Content Based Filtering is used toCollaborative FilteringCollaborative Filtering (short: CF) is the most commonapproach for Web 2.0 technologies. The simple comparisonof elements is extended by data on consumer behaviour. Sothe recommendation engine is able to predict items by characteristics of other users.In CF the user is related to items – here by a rating value.It is possible and more widely spread that not each user rateseach item. Actually some items may be never rated by a userand some users may never provide ratings at all. This issueis caused by the sparsity problem (cf. [24, p. 579]) and thecold start problem (cf. [7, p. 238]).In order to calculate recommendations the TV Predictormust know about the users interests. Therefore the enginediffers between two different feedback types (see figure 3): Automatically tracked watch behaviour: The clientsends in regular intervals messages indicating the watchedchannel, so the server can lookup for the current playing program in the database. Manually given ratings: The user can explicitly provide ratings for single programs in range of [1, 10].4.2.1Slope OneThe Slope One algorithm (Item-based Filtering) calculatesthe Top-N items by taking into account the ratings of allother users. It is divided into two parts. (cf. [20, p. 153])First of all there is a method to get the average deviation oftwo items:Pu Ui1 i2 (ru,i1 ru,i2 )dev(i1 , i2 ) (2) Ui1 i2 Where i1 and i2 are the items, ru,i1 is the items rating useru gave. Ui1 i2 is the set of users who rated both items and Ui1 i2 is its cardinality. So the result is the ratings deviationof an item. If the average rating of item i1 is higher than theaverage rating of item i2 the value is positive, if it is lowerthe value is negative, or if the average ratings are equal thevalue is dev(i1 , i2 ) 0.
The prediction value pre(u, j) for user u an
christopher.krauss@ fokus.fraunhofer.de Lars George Humboldt-Universität zu Berlin Unter den Linden 6 Berlin, Germany lars.george@ hu-berlin.de Stefan Arbanowski Fraunhofer Institute FOKUS Kaiserin-Augusta-Allee 31 Berlin, Germany stefan.arbanowski@ fokus.fraunhofer.de ABSTRACT Switching through the variety of available TV channels to nd the .
(predictor) is a good predictor of college GPA (criterion) The predictor variable is determined before the criterion variable, and the data span a length of time (4 years of high school) Multiple Regression Prediction Studies Suppose, in fact, that high school GPA is not the sole predictor
Reserved/Personalized License Plates-Application For VLIC-4.420 Original Date: 09/28/1990 Revision Date: 07/15/2017 Personalized Plate Policy Violations Create Personalized Message Relinquish Personalized/Reserved Plates Governor Series and Low Number License Plates Customer Requirements-Original/Reissue Front Counter CSR-Original/Reissue
ii. personalized medicine from a regulatory perspective 5 1. defining personalized medicine 5 2. fda's unique role and responsibilities in personalized medicine 11 iii. driving toward and responding to scientific advances 14 1. building the infrastructure to support personalized medicine 16 2. recent organizational efforts 20 iv.
Expanding the Frontiers of Personalized Medicine 9 PART III Evidence Development Despite the extraordinary pace of technological progress in personalized medicine, payers and providers rightly demand evidence demonstrating that personalized medical interventions can be integrated into health systems in ways that deliver both clinical and
"Personalized learning is a way to actually enact the pedagogy we believe in and that kids thrive in," said Tavenner, the founder of Sum-mit Public Schools, a California-based charter network that operates about a dozen of its own personalized-learning schools while licensing its personalized-learning software to hun-dreds of others.
Personalized Learning Visioning Tool Anyone interested in implementing personalized learning, including states, districts, schools, and individuals. Create consensus around the vision for personalized learning in your community; create draft communications to share with your community. Help create, refine, clarify, or communicate your vision for
Personalized learning. There is no single, universally agreed-upon definition of personalized learning (Pane, et al., 2017). That said, the. Aurora Institute's definition succinctly and powerfully incorporates many ideas common to various definitions: Personalized learning is "tailoring learning for each student's strengths, needs and
DISCOVERING STATISTICS USING SPSS PROFESSOR ANDY P FIELD 2 2. A regression predicting the mediator (Mate_Value) from the predictor variable (Age).3. A regression predicting the outcome (Gossip) from both the predictor variable (Age) and the mediator (Mate_Value).These models test the four conditions of mediation: (1) the predictor variable (Age) must
