The consumer market has been characterised by a high variety of products since the 1990's that have been ampilated by emergence of online traders and service-providers. For the objective of reduction of information overload that consumers are exposed to, stemming from the high degree of variety, new methods have emerged. The former is the reason underlying the attempts of recommender systems to utilize the outcomes and methods of the artificial intelligence.
My thesis aims at providing a comprehensive summary of the above scrutinised approach by evolving theoretical and practical insight respectively into the recommender systems.
First I review the theoretaical basis, which is applied by the methods used in recommender systems, in particular the intelligent solutions. In line with these I talk about the steps of data preprocessing and prediction methods. After that I examine, the way the above technics could be used by recommender systems, especially during the implementation of content based and collaborative filtering systems.
Based on the theoretical background I prepared an experimental prototype of an online movie recommender system. I review the important issues and planning decisions. The complete system use both content-based and collaborative filtering techniques to calculate recommendations, so that the two methods will be comparable through the example. In order to reduce the computation time for the recommendation, I used among others clustering techniques, but the way the compression of feature vectors were also touched upon. During the preparation I paid particular attention to the user interface usability and the attractive visualisation.
A real data set was used to the implementation, which comprises of millions of relational tables, hence is important to keep in mind the continuous optimization issues. The nature of the data set allow for the evaluation of the actual test results, stemming from the fact that contains existing user reviews.
At the and of paper I present a testing method, by the deployment of which the recommendation algorithm performance is measurable. To sum up the thesis I evaluate the system, highlighting its shortcomings and outline a few further development possibilities, such as model based recommendation.