Article Preview
TopIntroduction
Recommender systems have become extremely common in recent years, and are applied in a variety of applications such as music, news, research articles, search queries, social tags, and products in general. One of the goals of recommender systems is to help users navigating large amounts of data. Existing recommender systems are usually categorized into content-based methods and collaborative filtering methods (Adomavicius & Tuzhilin, 2005). Applying recommendation technology to multidimensional databases is an emerg- ing and promising topic, especially for recommending decision queries (Chatzopoulou, Eirinaki, & Polyzotis, 2009); (Stefanidis, Drosou, & Pitoura, 2009). It is of particular rel- evance to OLAP domain analysis which is inherently tedious since the user has to navigate through large data cubes to find valuable information.
OLAP systems users indeed formulate multidimensional queries to meet their specific needs for decision support. OLAP tools are known to be intuitive as their end users are not necessarily computer scientists. However, the large volume of data and the complexity of analytical queries which involve a lot of aggregations make this task of analysis more difficult to users. So it seems necessary to provide them solutions best suited to their way of thinking through methods of recommendation and enrichment of their analytical tasks. These methods are known under the name of personalization. In this paper, we propose a new personalization process of analytical queries to help users in their decision making. We are particularly interested in collaborative recommendation and enrichment of users’ decision queries based on their query log files.
Personalization research works which exploit query log files use in most cases fre-quent itemsets (Khemiri & Bentayeb, 2013) and association rules (Veloso, de Almeida, Gonçalves, & Meira, 2008). However, the large number of obtained frequent itemsets and association rules makes the task of personalization more difficult due to their volumetrics. Contrary to these approaches, the work we propose in this paper is based on another type of rules, more compact, called triadic association rules (Biedermann, 1997). These rules convey a richer semantics than conventional ones as they include a condition, in addition to the premise and the conclusion. Our personalization process is composed of five steps:
- 1.
Modeling users query log files of OLAP servers by a triadic context. This triadic context is composed of the users set, the queries set, and the attributes set (descriptors and measures) in the SELECT clause and a ternary relation between these three sets;
- 2.
Mapping triadic (tridimensional) context into a dyadic (bidimensional) one. This is done by flattening the set of users over the set of attributes;
- 3.
Computing dyadic association rules (premise → conclusion);
- 4.
Generating triadic association rules ((premise→ conclusion) (condition) through a factorization process of dyadic ones;
- 5.
Exploiting these triadic association rules for personalization.
To validate our collaborative recommender system in OLAP environment, we develop a software prototype P-TRIAR (Personalization based on TRIadic Association Rules). P- TRIAR extracts two types of rules from users query log files. The first ones will serve in query recommendation process by taking the collaborative aspect of users into account during their analysis sessions. This recommendation process will be carried out by the user communities discovered across multiple links between them. The second type of the extracted rules aims at enriching user queries by recommending to them relevant attributes that allow creating new decision queries. Preliminary experiments were conducted to as- sess the quality of the recommendations in term of precision and recall, as well as the performance of their on-line computation.