We presented ReDyAl which is a hybrid algorithm that dynamically uses both the transversal and hierarchical approach for discovering resources. It is independent of the application domain and, although we applied it to DBpedia, it could be easily adapted to other datasets in the Web of Data. It relies only on Linked Data and does not require to reduce the set of resources and links of the dataset to those belonging to a particular domain.
We evaluated and compared our algorithm against three state-of-the-art algo- rithms by conducting a user study and we also showed two practical applications of the algorithm by presenting a mobile application that provides movie recommen- dations and an eTourism mobile application, both relies on DBpedia. Although the algorithm could be applied to other datasets in the Web of Data, we selected DBpedia because it is a general dataset; thus cross-domain recommendations were possible. Besides, there is a high number of resources represented, a variety of do- mains addressed and it is continuously updated since it is extracted from Wikipedia. The user study demonstrated that ReDyAl improves in the novelty of the results discovered, although the accuracy of the algorithm is not the highest (due to its inherent complexity). Although ReDyAl is not bound to any particular domain, the study focused on movies because Telecom Italia was interested in a related use case. Furthermore, in this domain there is a quite large amount of data available on DBpedia and participants were not required to have specific skills.
Future work includes studying the relevance under different domains and improv- ing the accuracy of ReDyAl while maintaining its novelty. We plan to conduct other studies to compare it with traditional techniques and with approaches which combine Linked Data with traditional techniques. We are also working on combining ReDyAl with collaborative filtering techniques to take user preferences into account while providing recommendations. It is worth to note that ReDyAl could be extended to
consider more than one resource in input (e.g. all the resources rated positively by the user). In order to do this, ReDyAl could be executed multiple times to generate recommendations given a number of initial resources, and subsequently, the results could be merged. However, this would significantly increase the response time since the algorithm relies on SPARQL queries to discover candidate recommendations through the links among resources, which is computationally expensive. Thus, we should study how to do this taking performance into account. Another resource to consider could be the current context of the user. Context-awareness is addressed in Chapter6. In particular, in that chapter, we present a context-aware recommendation technique. We could also extend ReDyAl with the context-aware recommendation method presented in that chapter. For example, that method could be used to select an initial resource for ReDyAl from a set of user ratings based on the context.
Leveraging Ontologies for
Context-Aware Recommendations
6.1
Introduction
Context-Aware Recommender Systems (CARS) are a particular category of rec- ommender systems which exploits contextual information to provide more useful recommendations. For example, in a temporal context, vacation recommendations in winter should be very different from those provided in summer. Similarly, a restau- rant recommendation for a Saturday evening with your friends should be distinct from that suggested for a workday lunch with co-workers [1].
Nowadays contextual information such as time and location are easy to be obtained with modern devices. However, also other parameters may be considered, such as the company (alone, with friends, with the one’s partner) which may be relevant when recommending movies or vacations. In addition, the exact context sometimes can be too narrow, as Adomavicius and Tuzhilin [14] exemplified by considering the context of watching a movie with a girlfriend in a movie theater on Saturday. Using this exact context may be problematic for several reasons. First, certain aspects of the overly specific context may not be significant. For example, a user which watch a movie with the one’s partner in a theater may have the same preferences on Saturday and Sunday, but they may change on Wednesday. Therefore, it may be more appropriate to use a more general context specification, i.e. weekend instead of Saturday. Second, the exact context may not have enough data for accurate
rating prediction, which is known as the data sparsity problem. Thus it may be useful to refer to a more general context such as watching a movie with the one’s partner in a movie theater on the weekend, watching a movie with someone in a movie theater on the weekend, and so on.
Additionally, often user preferences and items representation depend on the application domain addressed or on the particular recommendation approach used. Thus, a significant effort is required to adapt the recommender system to another domain or to change the approach to use.
In this chapter, we address the problems previously mentioned and we focus on the following research questions:
• Is it possible to represent context by combining different dimensions (such as time, location, mood, etc.) and representing different granularities for each dimension (e.g. the precise time moment, the day of the week or the season)? • Is it possible to represent user preferences and items in such a way that can be adapted to different application domains and combined with different recommendation approaches?
We distinguish three forms of context-aware recommendation process: contextual pre-filtering, contextual post-filtering, and contextual modeling [14]. Pre-filtering approaches use the current context to select a relevant subset of data on which recommendation algorithm is applied. Post-filtering methods exploit contextual information to select only relevant recommendations returned by some algorithm. Contextual modelingdiffers from other techniques as it incorporates the context into recommendation algorithm. We opted for a pre-filtering strategy because it can be used with existing recommendation algorithms and avoids an expensive search of an effective post-filtering approach, as explained in Section6.2.1
We propose a new contextual pre-filtering approach which is based on two ontologies to represent context and user preferences: Recommender System Con- text (RSCtx)2 which describes the context, and Contextual Ontological User Pro- file (COUP), which represents user preferences. COUP is based on Structured- Interpretation Model (SIM) [90] and consists of multiple ontological modules. We
1An exhaustive review of CARS is out of the scope of this thesis. The reader may refer to the survey of Adomavicius and Tuzhilin [14].
evaluated our approach through an offline study with a rating prediction task which showed that the usage of the proposed ontologies and our pre-filtering technique with a number of well-known recommendation algorithms significantly improves the accuracy of prediction according to the Mean Absolute Error (MAE) measure.
The rest of the chapter is organized as follows: Section6.2provides an overview of CARS, Section6.3presents related work, Section6.4introduces our ontology to represent the context, while Section6.5addresses the overall recommendation approach and the representation of user preferences. We detail the evaluation process and its results in Section6.6and we conclude in Section6.7.