Article Preview
TopIntroduction
Current approaches for OLAP start by defining a multidimensional model (Rizzi, Abelló, Lechtenbörger, & Trujillo, 2006) in order to obtain a unified view with which all decision makers intuitively fulfill their information needs. However, this single multidimensional schema may be quite large and complex as it may deliver information to many kinds of decision makers who may have different needs (Stefanidis, Pitoura, & Vassiliadis, 2006), as we can see below in the motivation example. The main disadvantage is that decision makers are forced to understand and navigate the whole complex schema if they want to be able to find and acquire suitable data. To overcome these drawbacks, we propose a multidimensional schema that provides a personalized OLAP view for each individual user.
Figure 1 shows an overview of our approach for OLAP personalization: analysts interact by means of OLAP clients with an OLAP server. The proposed solution also includes a personalization engine which interacts with the OLAP client to provide the analyst with the personalized metadata and the OLAP server to manage the OLAP metadata. When defining automated personalization, the designer has to specify a set of rules that fulfill the system personalization requirements. It might happen that the personalization rules defined do not exactly fulfill the requirements, however since these rules are independent of the rest of the conceptual models and the implementation of the OLAP cube, they can be refined at execution time. The engine interacts with the analyst by means of the OLAP client providing her/him a personalized view of the data. It might happen that the final user wants to check all the data, with any filtering, which is always available. The personalization actions are only performed once during a session not to overwhelm the user with constant updates. This means that the adaptive information will be the same during the present session. This is the default option, but it is also possible that the designer decides when an adaptation should take place to fulfill each personalization requirement. Our approach can be described into two stages, first, we propose how to modeling the personalization in OLAP, and, in the second stage, we perform the generation of personalized OLAP metadata by applying a set of transformations within a model-driven strategy. Each of these stages is described in their corresponding section.
Figure 1. Overview of our approach for OLAP personalization
Due to the overwhelming volume of information that OLAP schemas contain, personalization is a key factor for improving the satisfaction of the decision maker (Stefanidis, Pitoura, & Vassiliadis, 2006). Several OLAP approaches provide decision makers with personalized views which focus on different aspects (see Related Work section for details). However, none of them permits the personalization of schemas at the conceptual level. This may cause several problems such as difficult maintenance, no independence of the target platform, evolution of the information requirements, etc. Furthermore, none of these approaches allow personalization to be applied during runtime (thus taking into account user behavior), but only preferences over data stated at design time. Moreover, current commercial OLAP tools (e.g., Oracle Discoverer, Pentaho Mondrian) take the whole multidimensional schema of the underlying data warehouse as input and allow designers to customize it by specifying which elements of a particular type are preferred to their peers. The main drawback of this is that designers have to customize the schema manually, and this process is prone to mistakes and is time consuming.
One of the benefits of our approach is that the information is obtained from the OLAP model in a customized manner according to the needs of decision makers. Therefore, while decision makers need to apply several OLAP operations to obtain the right data if the OLAP model is not personalized, the analysis is simplified since the personalized OLAP models only contain the required structures and the number of required operations is therefore lower.