Year of Graduation
A Recommender System Approach Based on Pattern Structures
School of Applied Mathematics and Information Science
This work examines potential applications of Patter Structures for use in tasks of providing recommendations. (Pattern structure is an extension of Formal Concept Analysis for working with data that have complex structure).The main focus of the research is on the design of movie recommender system based on Pattern Structures, and evaluation of its usefulness in practical cases.This work consists of three main parts. The first part deals with main definitions of Formal Concept Analysis, Pattern Structures and Interval vectors. (In our case Interval vectors are patterns).The second part of the research explores existing approaches to Recommender Systems. Furthermore, we explain the essence and illustrate applications of Slope One algorithm, which is used later for the evaluation of the results of this research. Finally, we introduce the algorithm for recommender systems based on Pattern Structures and provide an example of its usage.The third part of this research provides an overview of the data set we worked with and concepts of Precision and Recall measures. Besides, a description of software tools, implemented in MATLAB was given. We compared algorithms based on Pattern structures and on Slope One method in terms of Precision and Recall and find out the proposed method is better.Different papers and articles on recommender algorithms based on Formal Concept Analysis and on the earlier implementations of Pattern structures were studied to form a ground for this research.As a result of conducted experiments, Recommender Systems based on Pattern Structures and on Slope One method were implemented. For the experiments we used a wide-spread and freely available data set, MovieLens, which contains 100000 ratings given by 943 users to 1682 movies. Furthermore, a comparative study of recommendations quality of the latter Recommender Systems was done.The main result of this research is the development of the algorithms for Recommender Systems based on Pattern Structures. The result of these experiments shows that it provides good results in terms of precision and recall compared to Slope One baseline, thus it makes the method suitable for real-world applications.