RECOMMENDATION SYSTEM IN AN AUDIOVISUAL

DELIVERY PLATFORM

José Mª Quinteiro-González, Ernestina Martel-Jordán, Pablo Hernández-Morera

Aaron López-Rodríguez, Leidia Martel-Monagas

IUMA Sistemas de Información y Comunicaciones-Tecnología de la Información

Universidad de Las Palmas de Gran Canaria, Campus Universitario de Tafira, s/n, Las Palmas de Gran Canaria, Spain

Angelo Santana-del Pino

Departamento de Matemáticas, Universidad de Las Palmas de Gran Canaria, Campus Universitario de Tafira, s/n

Las Palmas de Gran Canaria, Spain

Keywords: Recommendation Systems, Ontology, Customization.

Abstract: One of the main tasks of the information services is to help users to find information that satisfies their

preferences reducing their search effort. Recommendation systems filter information and only show the

most preferred items. Ontologies are fundamental elements of the Semantic Web and have been exploited to

build more accurate and personalized recommendations by inferencing missing user preferences. With

catalogs changing continuously ontologies must be built autonomously without expert intervention. In this

paper we present an audiovisual recommendation engine which uses an enhanced ontology filtering

technique to recommend audiovisual content. Experimental results show that the improvements of the

ontology filtering technique generate accurate recommendations.

1 INTRODUCTION

With the Internet people are becoming overwhelmed

by information. Recommender systems seem to be

the choice to help users in the situation of filtering

information based on user preferences.

One of the most widely used recommendation

technique is collaborative filtering (CF) (Herlocker

et al., 1999). This technique recommends items

based on the experience of the user as well as other

similar users. Unfortunately, CF suffers from

sparsity and cold-start (Sarwar et al., 2001) (Vozalis

and Margaritis, 2003). Sparsity refers to the lack of

user ratings. This reduces the probability of finding

users who have rated the same items. Clustering

techniques partially solve these problems at the

expense of losing accuracy in the recommendations

(O´Connor and Herlocker, 1999). The cold-start

problem is a consequence of the unconstrained

search space which requires many items to be rated

in order to find the right neighbours.

Another recommendation technique is the

content-based filtering (Pazzani and Billsus, 2007).

where the user will be recommended with items

similar to the ones the user preferred in the past.

This kind of recommendation technique does not

suffer from the cold-start problem. However, this

recommendation technique always ends up

recommending items which have already been rated

by the user (overspecialization).

An alternative to solve the above-mentioned

problems is to use the ontology filtering technique

(Schickel-Zuber, 2007). This technique infers user

ratings using a mechanism over an ontology of the

domain.

In this paper we use the ontology filtering

technique on an audiovisual delivery platform. This

recommendation technique has been exploited to

customize both audiovisual content and

advertisements according to the user profile and

their behavior in the platform.

379

Quinteiro-González J., Martel-Jordán E., Hernández-Morera P., López-Rodríguez A., Martel-Monagas L. and Santana-del Pino A..

RECOMMENDATION SYSTEM IN AN AUDIOVISUAL DELIVERY PLATFORM.

DOI: 10.5220/0003063103790382

In Proceedings of the International Conference on Knowledge Discovery and Information Retrieval (KDIR-2010), pages 379-382

ISBN: 978-989-8425-28-7

 2010 SCITEPRESS (Science and Technology Publications, Lda.)

2 BACKGROUND

In this section we present the scenario where the

recommendation engine has been developed, and the

ontology filtering technique.

2.1 Audiovisual Platform

The audiovisual platform architecture is basically

formed by repositories, managers and a user

interface. This architecture follows the SOA

(Services Oriented Architecture) model where each

manager provides its functionality through web

services.

The user manager module is in charge of user

authentication as well as the maintenance of user

session and identification through the platform.

The content manager module manages the

audiovisual content resources.

The virtual character manager manages the

appearance and behavior of virtual character that

interact with the user.

The statistic manager module is in charge of

maintaining the key performance indicators.

The customization manager provides the most

adequate audiovisual contents to users regarding

their taste and preferences.

The advertisement manager is in charge of

managing all the data regarding advertisements

within the audiovisual platform.

2.2 Ontology Filtering

Ontology filtering (Schickel-Zuber, 2007) solves the

problems of other recommendations techniques by

means of a ontology containing the item domain,

and an inference mechanism which allows inferring

user preferences using the proximity to other items

of the ontology which have been rated by the user.

A node in the ontology represents a concept. An

item is an instance of a concept, and the edges

represent features that differentiate the set of sub-

concepts of the same parent.

For each concept c, two properties are added: the

score (S(c)) and the a-priori score (APS(c)). The

former represents how much a specific user likes a

given concept, whereas the latter determines the

propagation parameters of user’s scores between the

sub-concepts and the concept c, considering only the

location of concept c in the ontology.

To predict user ratings of a concept y, the system

uses knowledge of the user ratings, S(x) associated

to the closest concept x, and the lowest common

ancestor to concepts x and y in the ontology.

Recommendations are computed by finding

unseen items that are instances of concepts with the

highest score.

Changes in the set of items are incorporated in

the catalog, through the use of clustering algorithms

to generate ontologies from the set of user ratings.

The clustering algorithms used are the

agglomerative and the partitional. The criteria

functions used to optimize the allocation of an item

to a given cluster are shown in Table 1.

Table 1: Criteria functions used by clustering algorithms.

where k is the amount of cluster to consider, C

represents cluster r, n

denotes the amount of

elements in cluster r, C

is the centroid of cluster r,

and sim(i,j) is the similarity between items i and j.

Combining the criteria functions in Table 1 and

the clustering algorithms, the system generates 15

different ontologies (agglomerative with criteria

function 1 to 9, partitional with criteria function 1 to

6). Instead of using the same ontology for all users,

the best ontology is selected based on the user

preference profile in order to increase the

recommendation accuracy.

3 RECOMMENDATION ENGINE

ARCHITECTURE

The recommendation engine has a layered

architecture, where the inner layer contains the

agents that implement different filtering techniques.

Above this layer, the recommendation agent drivers

let you configure what techniques to use and how to

combine them, and to determine the set of item

ratings, product catalogs and ontologies to be used.

The architecture allows us to add new

personalization techniques easily by implementing

the corresponding agent. This goal has been

achieved by means of software architecture

presented in Figure 1. The decision tree chooses the

recommendation technique which is carried out by

KDIR 2010 - International Conference on Knowledge Discovery and Information Retrieval

380

Figure 2: Class hierarchy for recommender.

the recommender. The recommender sets up the

interaction between agents and data management.

This interaction allows agents to get the required

data for providing recommnedation results.

Figure 1: Software architecture for the application layer.

The class hierarchy which implements the

recommender is presented in Figure 2 where

filtering techniques have been separated from the

data required for recommendations. Each new agent

requires adding a new class for the recommender,

another for the filtering technique and the necessary

data classes for such a filtering technique.

The agent that implements the ontology filtering

technique has two components: the ontology

generator and the recommendation generator. The

ontology generator computes off-line the set of

ontologies specified in subsection 2.2. The

recommendation generator selects the most

appropriate ontology for each user.

Based on the selected ontology, the user ratings

and the product catalog, the recommendation

generator performs the necessary tasks to generate

the recommendations.

4 EXPERIMENTAL RESULTS

We have been working with an improved version of

the ontology filtering technique, by estimating the

missing values in the user-item matrix by means of

an imputation method. The algorithm of the k

nearest neighbours has been used as the imputation

method of missing values (Troyanskaya et al.,

2001).

To evaluate the recommendation engine with knn

imputation values, several experiments have been

performed using the data from the MovieLens data

set (Dataset MovieLens, 2010). 250 users from this

dataset have been randomly selected.

The set of ratings of each user is randomly split

into a training set and a test set, so that the former

contains 90% of the ratings.

The training set is used to generate the

ontologies. 15 ontologies have been generated by

combining agglomerative and partitional methods

with each criteria function listed in Table 1.

Similarly, another 15 ontologies have been

generated but from an array of user-item ratings with

data imputation. All of this leads to a total of 30

ontologies.

For each of the ontologies generated, the inferred

scores are compared with the real scores of those

concepts in the test set, computing the NMAE

(Normalized Mean Absolute Error) (Herlocker et al.,

2004).

Then, the ontology with the lowest average error

is selected. The results are shown in Table 2, which

shows the number of times a clustering method with

a specific criteria function has been selected.

Table 2 shows that more than half, namely 134

users are benefiting from the use of knn data

imputation.

The mean error using the ontology filtering

method is 0.056, and in the case of using ontology

filtering with knn data imputation the error is 0.048.

RECOMMENDATION SYSTEM IN AN AUDIOVISUAL DELIVERY PLATFORM

381

Table 2: Distribution of selected ontology for each user.

Clustering method

agglom. agglom-knn part. part-knn

Criteria function

1 13 16 12 0

2 6 8 4 4

3 4 9 3 2

4 9 3 0 11

5 11 14 9 5

6 3 8 2 2

7 22 12 NA NA

8 8 25 NA NA

9 10 15 NA NA

Total

86 110 30 24

In relative terms this means an improvement of

14.1%.

To determine if this difference between the

average errors of both methods is significant we

have used the Wilcoxon test for paired samples,

since there is no normal distribution of errors, as can

be seen in Figure 3.

Figure 3: Wilcoxon test.

The p-value Wilcoxon test is less than 10

-6

which

indicates that the difference between the magnitudes

of the errors is significant and cannot be attributed to

chance.

5 CONCLUSIONS

In this work we have presented the recommendation

engine developed for the customization and

advertisement managers within an audiovisual

content delivery platform. The main novelty of our

work is the reusable approach of the

recommendation engine for different contexts. In

this sense, we can add new different filtering

techniques and different content types easily. In this

work we have used the ontology filtering technique.

This work has enhanced the traditional ontology

filtering technique by means of an imputation of

ratings in the user-item matrix.

6 FUTURE WORK

The extended use of social networks introduces an

important element to extend user profile as the user

interacts with other users in the network. Moreover,

some geolocalization aspects must be considered for

recommending advertisements.

At the moment we are updating the ontology

filtering technique in order to improve the quality of

the recommendations even more.

REFERENCES

Herlocker, J., Konstan, J.A., Borchers, A, Riedl, J., 1999.

An Algorithmic FrameWork for Performing

Collaborative Filtering. In 22nd Annual International

ACM SIGIR Conference on Research and

Development in Information Retrieval, Berkeley, CA,

pp. 230-237.

Sarwar, B., Karypis, G., Konstan, J., Riedl, J. 2001. Item-

based collaborative filtering recommendation

algorithms. In 10th Conference on World Wide Web,

Hong Kong, pp. 285-295.

Vozalis, E., Margaritis, K.G. 2003. Analysis of

Recommender Systems’ Algorithms. In 6

Hellenic

European Conference on Computer Mathematics and

its Applications (HERCMA), Athens, Greece.

O´Connor, M., Herlocker, J. 1999. Clustering Items for

Collaborative Filtering. In Proceedings of the ACM

SIGIR Workshop on Recommender Systems:

Algorithms and Evaluation, Berkeley, CA.

Pazzani, M.J., Billsus, D. 2007. Content-Based

Recommendation Systems. Lectures Notes in

Computer Science, vol. 4321, Springer, pp. 325-341.

Schickel-Zuber, V. 2007. Ontology Filtering. Thesis,

Lausanne, EPFL.

Troyanskaya, O., Cantor, M., Sherlock, G., Brown, P.,

Hastie, T., Tibshirani, R., Botstein, D., Altman, R.B.

2001. Missing value estimation methods for DNA

microarrays. Bioinformatics, vol. 17, n. 6, pp. 520-

525.

Dataset Movielens. http://www.cs.umn.edu/

Research/GroupLens/data (visited March 23, 2010).

Herlocker, J.L., Konstan, J.A., Terveen, L.G., Riedl, J.T.

2004. Evaluating Collaborative Filtering

Recommender Systems. In ACM Transactions on

Information Systems, vol. 22, n. 1, pp. 5-53.

KDIR 2010 - International Conference on Knowledge Discovery and Information Retrieval

382