Conventional recommender systems provide personalized recommendations by collecting and retaining user data, relying on a centralized architecture. Hence, user privacy is undermined by the volume of information required to support the personalized experience. In this work, we propose a User Experience model which allows the privacy of a user to be preserved by means of a decentralized architecture, enabling the Service Provider to ofer recommendations without the need of storing individual user data. We advance the current state of the art by: i) Proposing a model of User Experience (UEx) suitable for Persona-based recommendations; ii) Presenting a UEx collection model which enhances the user privacy towards the service provider while keeping the quality of her preferences predictions; and iii) Assessing the existence of the Persona profiles, which are needed for generating and addressing the recommendations. We perform several experiments using a real-world complete dataset from a medium-sized service provider, composed of more than 14,000 unique users and 33,000 content titles collected over a period of two years. We show that our architecture, in combination with our UEx model, achieves the same or better results, compared to state-of-the-art systems, in terms of rating prediction accuracy, without sacrificing user's privacy.
In the early days, broadcast television was a one-to-many relationship. The signal traveled one-way from the content provider towards the consumer, and so did the contents. User privacy was Copyright © CIKM 2018 for the individual papers by the papers' authors. Copyright © CIKM 2018 for the volume as a collection by its editors. This volume and its papers are published under at the highest possible level, since the user was anonymously and passively connected to the network.
With the introduction of an Internet-based "return path" for
voting or rating of programs, initially in Digital Video Broadcast
set-top-boxes, and later with Over-the-Top (OTT) TV [
By removing the constraint of one-way communication and
assigning a unique address to each node of the network, service
providers can easily collect detailed information from each node to
build user profiles, and the level of personalization of the service
theoretically has no limit [
Many countries are pushing towards the concepts of privacy by
design and privacy by default, in particular in the European Union
(EU), driven by the General Data Protection Regulation (GDPR) law,
which strongly highlights the concepts of linkability and personal
data [
Inspired by these GDPR principles, and focusing on a solution
that could ethically improve the “status quo”, we here present the
following contributions:
(1) We design a mathematical User Experience Model that allows a
Service Provider to collect User Experience in a privacy-aware
system that keeps personal information under the user’s control within
her domain and a sanitized database within the Service Provider’s
domain.
(2) We validate the User Experience Model, which is used for user
classification, by computing a type of rating based on weighted
predictions with a many-fold comparative experiment, testing various
similarity metrics and prediction algorithms.
(3) We perform extensive experiments using a real-world full
dataset from a medium-sized service provider, composed of more
than 14,000 unique users and 33,000 content titles collected over a
period of two years. We use the Root Mean Squared Error (RMSE)
to compare the performance of our solution with state of the art
algorithms such as the FunkSVD, which is the winner of the popular
Netflix competition, ItemItem and PearsMean [
The results show that our User Experience Model achieves the
same or better results, compared to state-of-the-art systems, while
ofering the potential for drastically increasing the user privacy.
It enables Service Providers to accurately classify users’ tastes by
storing only a fraction of the users consumption data and thereby
reducing drastically the linkability hazard. Moreover, we show how
clustering techniques applied to the User Experience Model of
an OTT TV user base leads to the description of distinct Persona
profiles, defined as homogeneous groups of individuals whose
consumption patterns and motivation can be represented as a set of
statements derived from quantitative measures [
The concept of Persona is necessary for the anonymous user classification according to Persona profiles. Service personalization can then be made using the Persona profile to which a user belongs when accessing the service.
Our solution aims at assisting providers in fulfilling the obligations enforced by law. A privacy-enhanced design of the recommender system can provide a competitive advantage in at least two aspects for providers: i) Lower expected costs for implementing future system updates in order to comply with new regulatory restrictions; and ii) Lower risks for privacy litigations, such as the case settled by Netflix for $9 million in favor of its customers 1. 2
In this section, we present the related work of the privacy and context-aware Recommender Systems, providing also the technological and regulatory background of the privacy problem which we define as follows:
How can we enable OTT TV providers to drastically enhance their users privacy by augmenting existing technologies, protocols as well as recommender systems? 1Case No. 11-cv-00379 (U.S. District Court for the Northern District of California) The intuition behind our solution is that a provider can satisfy the privacy of individual users by providing recommendations to groups of users with similar characteristics. The solution is based on the concept of Persona and Locked Persona Profiles . Succeeding in solving this challenge is not only beneficial for the users, but also convenient for the service providers. We provide a solution by breaking such a problem into the following research questions: • How to turn the passive collection of consumption data identified by the ontology User <-> consumes <-> content, into a model of User Experience (UEx)? • How to collect consumption patterns while avoiding the linkability to the related personal records? • How to provide Persona profiles from such a model of User
Experience? • Do Persona profiles exist in the OTT TV context? 2.1
Digital Rights Management (DRM) technology has been specified,
designed, and implemented in order to fight piracy. As such,
satisfying DRM sets a minimum level of possible privacy for the user.
Yang et al. [
In order to provide a personalized recommendation to any
anonymous user accessing DRM protected contents, the design approach
theorized by Cooper et al. [
In [
A Recommender System usually needs to collect some amount
of personal information about the user in order to provide the
recommendation. This includes attributes, preferences, contact lists,
among others [
Several methods have been proposed in the past for reducing
linkability – here defined as the possibility of discovering a
relationship between a user’s consumption records or between a user’s
consumption and her identity [
User Experience (UEx) is very important for the service providers. Considering the specificity of the OTT TV application field, the User Experience should be exploited to provide users with good recommendations about linear or Video on Demand (VoD) media contents. Therefore, the interest is not the User Experience for OTT TV service itself, but rather the User Experience about the content available in the Electronic Program Guide (EPG).
As such, service providers define several metrics related to User
Experience:
Service Consumption. In order to have any experience with the
media content, users need to consume the content [
Subjective Perceived Value. A positive or negative experience
results from the contribution of multiple drivers [
According to the review presented in [
The main solutions described in the literature to preserve
privacy are the following. (1) “Privacy preserving approach based on
peer to peer techniques using users’ communities” with
recommendations generated on the client side without involving the server [
The first step towards achieving the architecture described in the previous section is to understand who is consuming what, when, where, and whether she liked what has been consumed (or not). Our dataset consists of more than 700,000 ZAP events, representing the fact that a user taps on a content, 14,000 unique users, and 33,000 content titles collected over a period of two years by a medium-size OTT TV provider.
We define the User Experience Model (UEx) as a tensor: U Ex : C1m1 ⊗ C2m2 ⊗ · · · ⊗ Ckmk ⊗ Pmp 7→ IRm1+m2+···+mk +mp (1) where each of the Cxmy represents a vectorial space of the xth context in which the user zaps into media contents, and Pmp represents the vectorial space of the media contents that she zaps; m1, m2, · · · , mk , mp are the dimensions of each vectorial space.
In a traditional recommender system users may provide ratings,
and Latent Semantic Analysis (LSA) [
Linearizing the tensor defined in Eq. (1), we obtain the User Experience matrix presented in Eq. (2). The matrix consists of the Zap Frequency Inverse Context Frequency, one slice for each context, and the Zap Frequency Inverse Program Frequency (last slice). Each slice, except the last one, represents one of the k possible contexts of consumption, while the last slice represents the TV programs
U Ex(1) Z F ICF1,1,1 · · · Z F ICF1,1,m1 · · · © U Ex(2) ®ª® = © Z F ICF2,1,1 · · · Z F ICF2,1,m1 · · · ... ®® ... . . . ... ... « U Ex(n) ¬ « Z F ICFn,1,1 · · · Z F ICFn,1,m1 · · ·
· · · · · · . . .
Z F ICFn,k,1 · · · Z F ICFn,k,mk
Equation 2: Linearized representation of the tensor described in Eq. (1). It is composed of slices corresponding to each of the ifve features, i.e. the four contexts (i) Time of Day, (ii) Time of Week, (iii) Time of Month, and (iv) Time of Year, see Eq. (3), plus one (v) concerning the TV-Programs consumed by the user, see Eq. (4). Each row maps the ith user experience UEx(i). Therefore, just by looking at the distances between the users represented, this model allows measuring the similarity between their experiences against the TV-Program consumed within the contexts. consumed. Within each slice, except the last one, each column represents the mth split of the kth context, while each column of the
k last slice represents one of the TV titles (or programs).
Each row represents the experience of the ith user in the time span of her interest. However, it could also represent the experience of the ith component of a cluster (Persona) in the time span of Service Provider interest. Therefore, the experience of each user can be collected as a vector having the same representation of one row of the matrix. This vector would be the summary of the User Experience and could be easily collected by the Service Provider and be put under the user’s control by using OAuth 2.0 on the zapsto-contents. Besides, if the users mapped in this way are arranged in clusters, each cluster would represent homogeneous experiences.
If such homogeneous groups exist, they allow for user classification as well as a “set of statements” derived from the measures, fitting to the definition of a Persona.
In this subsection we elaborate on the meaning and the nature of the User Experience components, and discuss the concepts of Z F ICF and Z F I P F . 3.2.1 Zap Frequency Inverse Context Frequency. The Zap Frequency Inverse Context Frequency relates to the first m1 + m2 + · · · + mk components of the User Experience tensor defined in Eq. (1). We define the Zap Frequency Inverse Context Frequency (ZFICF) for user i on the context segment j of context k as:
Z F ICFi jk =
Zi jk T ot Zik · log T ot Ck , where:
ACi jk (3) • i indicates the ith user, where i = 1, ..., n; • the kth context can refer alternatively to Time of Day, Time of Week, Time of Month, Time of Year, or Device (D), i.e. k = 1, ..., 5. • j indicates the jth segment of the kth context, j = 1, ..., mk .
The kth context is divided into a total number of segments equal to T ot Ck = mk . For example, Time of Day could be divided in 24 segments, thus mT imeof Day = 24, one per hour; but it could also be divided in 2 segments as day and night, in such case mT imeof Day = 2.
The first factor of Eq. (3) is the Zap Frequency (ZF) recorded for the user i in segment j of context, where Zi jk represents the amount of zaps recorded during the time span of interest, and T ot Zik = ACik = card j | Zi jk , 0, ∀j ϵ [1, mk ] Ímj=k1 Zi jk is the total amount of zaps recorded for user i in context k during the (same) time span.
The second factor is the Inverse Context Frequency (ICF) recorded for user i in context k, where T ot Ck = mk , as already mentioned, is the total number of segments composing the context k; and indicates the number of segments within the context k, in which the ith user zapped at least once. 3.2.2 Zap Frequency Inverse Program Frequency. The Zap Frequency Inverse Program Frequency (ZFIPF) relates to the last mp components of the User Experience tensor defined in Eq. 1. It is similarly defined by the following equation:
Zi j Z F I P Fi j = T ot Zi · log APi
T ot P , where: (4) • i denotes the ith user, where i = 1, ..., n; • the jth element indicates the jth TV program, where j = 1, ..., mp . The total amount of TV programs available on the EPG is T ot P = mp .
The first factor of Eq. (4) is the Zap Frequency (ZF) for the ith user on the jth program, where Zi j represents the amount of zaps recorded during the time span of interest for user i on the TV program j, and T ot Zi = Ímj=p1 Zi j is the total amount of zaps of user i within the (same) time span.
The second factor represents the Inverse Program Frequency (I P Fi ) for user i in the time span of her interest, where T ot P is the total amount of TV programs available on the EPG, and APi = card j | Zi j , 0, ∀j ϵ 1, mp indicates the number of programs for which user i zapped at least once within the (same) timespan.
It is important to note that the ZFIPF over a large interval of time relies on the grouping of the items, especially in environment where, for example, TV series or news programs might recur continuously. If the grouping is not done properly, then the measure will be afected by a large error.
In this section, we provide several experiments that prove the
existence of Persona profiles based on User Experience. They can be
efectively computed using a Zero-Knowledge Consumption Data
Base (DB), which we define as a collection of data sampled on-line
from the user consumption feedback in such a way that only a
portion having size below 10% of the consumption data is retained
and stored in the DB in order to fulfill the definition provided for
Zero-Knowledge in [
As the results of these experiments, we show and validate that: i) User Experience is efective for classifying a user according to such Persona profiles; ii) Zap Frequency Inverse Program Frequency predictions computed by looking for similar users from a sample of the dataset are reasonably accurate;. Finally, we provide an example of Persona profile, obtained by clustering a sample of the dataset, according to the definition stated in the introduction.
We use a real-world dataset, provided by a Media Company, which provides both Video On Demand and Linear TV on behalf of a Content Provider2. The data can be represented in a minimalistic and generic way by the following: <contractId, contentId, deviceId, deviceType, zap_timestamp>, where: contractId is the ID assigned to the contract by the Content Provider, e.g. at the start of subscription, in order to authenticate the user, referring to one or possibly more users active with the same contract; contentId is the ID assigned to the media content by the Content Provider, e.g. the TV Network; deviceId is the ID of the user’s device authenticated while accessing the content, this is also used in combination with the contractId to identify a single user (userId); deviceType identifies e.g. Mobile, PC or TV; and zap_timestamp is the timestamp determining when the user switches the TV channel (it is an attribute of the ontology user - zaps into - content).
At this stage, for the purpose of organizing users in homogeneous groups, it is necessary selecting data about any user consuming contents alone and filter out data about any user consuming contents together with other individuals, because while the first group represents a clean signal we want to analyze, the second is afected by noise. Therefore, since it seems reasonable to consider mobile and tablet devices as personal and separate them from PCs and TVs, which are considered social devices. Distinguishing these devices based on the screen size3, from the full dataset we restrict the data only to mobile devices and only keep the following information: <userId, contentId, contentTitle, zap_timestamp>. The resulting dataset is composed of 743975 Zaps4, 14518 Users, and 33357 Content titles consumed over 2 years. 4.1
The purpose of this experiment is to verify the consistency of the
User Experience modeled in Eq. 1 by predicting 90% of the users’
preferences employing only 10% of them for the computation.
Succeeding in such a challenging purpose will demonstrate that Service
Providers could serve recommendations by maintaining only 10%
of the data currently retained about the users consumption. The
experiment consists in two steps: i) Testing the consistency of
our solution by predicting the Zap Frequency Inverse Program
Frequency (ZFIPF) and measuring the RMSE. We use 90% of the dataset
to find target users and 10% for computing the prediction. ii) Testing
the consistency of the User Experience Model by applying three
2Data provided by Mediathand, company based in Denmark, which operates the OTT
TV service on behalf of Glenten.
3We assume that the chance of a user watching some content together with others is
directly proportional to the screen size.
4Event recorded when a user taps into a content.
popular algorithms implemented within the last updated version
of Lenskit [
Moreover, in order to avoid the Cold Start Problem (CSP) related
to Users [
The User is represented by her User Experience Vector which is
a row of the matrix in Eq. 2. Thus, we classify each User belonging
to the Target Users Sample represented as in Eq. 2 against each of
the relevant users belonging to Model Users Sample represented as
in Eq. 2. Each random sample counts 10% of the population from
which it has been extracted [
The user classification and the ZFIPF prediction have been
accomplished by the following steps. i) Compute the User
Experience vector for each user of both the Target Users Sample and the
Model User Sample. ii) Store the Zap Frequency Inverse Program
Frequency (ZFIPF) regarding each media content within the TCS
from the User Experience Vector (UEV) of each user belonging
to the Target Users Sample in a new array (NA) and then set to
zero the ZFIPF within the UEV origin. iii) For each user belonging
to Target Users Sample, compute the K-Nearest Neighbor
belonging to Model Users Sample having the relevant Zap Frequency
Inverse Program Frequency not-null, using all the mentioned
metrics. iv) For each metric, compute a Zap Frequency Inverse Program
Frequency as mean of the values available from KNN. v) Compute
the Root Mean Squared Error (RMSE) of each prediction from the
NA. The random prediction has been generated using the range
Z F I P F ∈ [0, max (NA)] instead of Z F I P F ∈ [0, ∞], therefore it is
more accurate.
4.1.2 Lenskit Experiment. Lenskit is a very popular and efective
tool developed by the Grouplens from the Minnesota University [
PearsMean
Item
FunkSVD
NN-MEAN SVDcos NN-MEAN SVDcheb RANDOM It encapsulates several recommender systems algorithms such as ItemItem, PearsMean and FunkSVD. In particular, these have been taken into account for this experiment. The experiment was run on the ZFIPF computed using the whole dataset and provided in the same format of the Movielens ratings. The setup of the experiment is pretty straightforward and involves the selection of the ratings domain, which we set in the range Z F I P F ∈ [0, max (NA)]. The precision has been set as maximum resulting from ZFIPF. Through some preliminary experiment we found the parameters ensuring the lowest RMSE. For example, the amount of features for FunkSVD, which is set to 40 by default, performs very poorly, while at 10 seems to achieve the best performance. Lenskit completes a 5-fold validation on the whole dataset. The RMSE results of the experiment are around 0, 4 and are presented in Figure 1. 4.1.3 Experiment Outcome (Figure 1). We aim at proving the consistency of the User Experience Model by comparing RMSE resulting from the experiment run with Lensikit against the results harvested using our solution. From the cross-validation based on the same dataset, it is possible to notice as a first result that the model predictions outperform the random assignment and are aligned to the state of the art algorithms provided by Lenskit. Moreover, from these results we can notice that the similarity metrics used for classification do not influence the RMSE, they all yield closely the same results. However, we chose Chebyshev, because e.g. compared to Cosine it could identify clusters where users are closer to a normal distribution. 4.2
This experiment aims to finding distinct Persona Profiles. Using mostly the same set up of the previous experiments, therefore, a sample of users having 10% of the whole population’s size, this experiment should take into account the following challenges: i) Which clustering algorithm is the best fit for finding distinct groups of similar users? ii) How to choose the amount of clusters in order to avoid over-fitting?
To choose between k-means [
Our analysis and experiments have shed light upon significant additions to the solution of the problem as formulated and the following statements are an attempt of summarizing these contributions: (a) It is possible to provide Persona-based recommendations for OTT TV, because apparently homogeneous group of users exist. (b) Based Name Size Persona profile Clust. 6 59% Addicted to News Content . Mainly active during the EveningToD, Mostly during the week, never SaturdayToW and slightly more during SummerToY . Clust. 2 17,6% News is first, then fitness Content . Mainly active during Evening (less at Night and Noon)ToD, Mostly during week, never SaturdayToW . Clust. 9 5,9% sPlaigsshitolynamteoarebofaurt fSrpoomrtPNayewCsh, eTcaklTenotMS,hSouwmamnedrCanards,AIunttuermenstTeodYin Crime, some times cartoons (perhaps for intertaining her children)Content . Clust. 1 4,5% PMaasisniolynaatcetiavbeoduutrCinygcltihnegEavnednIinntger(nesetveedriant FNoiogthbta)TlloCDonMteonsttl.yMdauirnilnygatchteivweedeukr,insgomMeotrinminesg ianntdhEevweeneinkgen(sdoTmoeWt,i mfaersfraotmNiPgahyt)TCohDec,kFTaro MfroanmdPdauyrcinhgecWkTionMtearTnodYSummerToY . Clust. 7 4% Passionate about Cooking and Drama, Interested in News and CrimeContent . Mainly active during Morning and Evening (never at Noon)ToD.
Mostly during the week, never SaturdayToW , Far from Pay CheckToM and AutumnToY Clust. 8 2,25% Passionate about Fitnes and Interested in Sport News and NewsContent . Mainly active at Noon, never at NightToD. Mostly Active in Winter, Never during Spring.ToY . Clust. 3, 4, 5, 10, 12, 13 < 1, 8% Seed Clusters Algorithm 1: It samples from the dataset and it computes the User Experience matrix (see Eq. 2). Then, after dimensionality reduction using Singular Value Decomposition, it estimates the amount of clusters and it detects the clusters using hierarchical clustering. Finally, for each cluster it returns the measures on the features necessary for providing the Persona profiles. 1 function DetectPersonaProfiles ;
Input : Dat aset , M ZT = 25 , S ampl inдCoe f f icient = 0.1 ,
M eanMin = 0.05 , S t DevMax = 1 , Simil ar ityM et r ic = Chebyshev , Clust er s Est imat ionM et hod = Dav Bou ,
Clust er inдT ech = H ier ar chicalClust er inд; Output : P er sona(Clust er Size , T O PCont ent s List , mean(Z F I C FT oD ) , mean(Z F I C FT oW ) , mean(Z F I C FT oM ) , mean(Z F I C FT oY )) 2 MUS = Sampling(M ZT , S ampl inдCoe f f icient, Dat aset ); 3 Zaps = CollectZaps( Dataset, MUS); 4 Z F I C Fi jk = zficf(Zaps); 5 Z F I P Fip = zfipf(Zaps); 6 User Experience =
H or izont alConcat enat e (Z F I C Fi jk, Z F I P Fip ); 7 [U,V,S] = SingularValueDecomposition(User Experience); 8 ClustersAmount = ClustersEstimationMethod(U, SimilarityMetric); 9 Clusters = ClusteringTech(U, ClustersAmount, SimilarityMetric); 10 for Cluster ∈ Clusters do 11 return Clust er Size = 100∗ccoouunnttEElleemmeenntt((MClUuSst) er ) ,
T opCont ent s List = {Cont ent s | mean(Z F I P FCont ent ) > M eanMin & S t Dev (Z F I P FCont ent ) < S t DevMax }, mean(Z F I C FT oD ), mean(Z F I C FT oW ), mean(Z F I C FT oM ), mean(Z F I C FT oY ); 12 end on the ontology represented by the dataset, it is possible to define a model of User Experience that can be collected and updated by the Service Provider in samples (as e.g. the Model Users Sample) which no longer would be linkable to users. Therefore, it would be possible for the Service Provider to collect a privacy-aware Zero-Knowledge Consumption DB. In particular: i) We defined a mathematical representation of User Experience suitable for the specific application ifeld, which has been implemented and tested thoroughly with dimensionality reduction and hierarchical clustering. ii) We showed that, in this application field, Personalization-based on Persona-profiles is possible, homogeneous groups of users exist, and familiar clustering techniques can distinguish between users with diferent interests even when close to each other, such as those preferring sports Vs. fitness. iii) We demonstrate that the quality of prediction of a popular Recommender Systems such as FunkSVD, where the user privacy represents a huge issue, is comparable with our predictions, where we potentially can keep the privacy at ZK-level since we need only 10% of the user data constantly stored on the service provider side and this would make extremely dificult linking any user to such data set.
Further work should also be prioritized towards the following focus
areas. i) Data sampling. It is a very critical component of both
Recommender System and Privacy enhanced architecture. More
sophisticated and efective options, such as the smart sampling
presented by Google [
The authors would like to acknowledge the help provided by Hossain Ahmad and Naoufal Medouri. 0.15 -0.2
-0.1 -0.15 -0.05