Effect of refactoring on the programmers' workload


Tezin Türü: Yüksek Lisans

Tezin Yürütüldüğü Kurum: Orta Doğu Teknik Üniversitesi, Mühendislik Fakültesi, Bilgisayar Mühendisliği Bölümü, Türkiye

Tezin Onay Tarihi: 2013

Öğrenci: HÜSEYİN CAN DOĞAN

Danışman: ALİ HİKMET DOĞRU

Özet:

Recommender systems are very popular in information systems and in the research community, where many different approaches geared towards giving better recommendations have been proposed. In this thesis, we propose a methodology that uses social network information to improve the performance of recommender systems. Our proposed methodology heuristically improves the success rate and performance of recommendation algorithms using social distance measures on a dataset that comprises people in professional occupations. Further, we explain how these methods apply to on-line real-world applications. The main objective behind the composition is to provide better and more relevant inputs to item-to-item filtering algorithms. We propose a compound method comprising three steps. In the first step, the algorithm elaborates social network distances and friendships to help recommender systems customize the target user set. To find people who are similar to a specific user, the system divides each worker's friends (target set) into subsets and treats the task as a social clustering problem. In the second step, clustering is done on social measures. The clustering algorithm divides the target set into subsets to build a job-to-job table and a similar-job pairs of people who tend to do the same kind of work. In the third step, highly recommended jobs are defined by computing distance metrics on job vectors. Thereby, item-to-item recommendation can compute ordered predictions for users. We interpret the differences between social-based relations and the impact of similarity metrics on a collaborative recommendation algorithm. The experiments conducted on large datasets indicate that our proposed approach, which customizes recommendations using social connections, outperforms generic methods in terms of specificity and scalability. We also conducted several experiments to compare the evaluation and recommendation qualities of our approaches with other well-known algorithms such as Restricted Boltzmann Machines. Our evaluations show that the components of our method combine to facilitate deeper understanding of the performance characteristics of recommender systems.