Comparison of machine learning techniques used in the recommender system models in accordance with customer purchase preferences

Recommender systems in e-commerce are the collection of models that aim to reveal products that customers would prefer after evaluating product features via information obtained from customers. In the recommender systems, the efforts made in order to guess the objects which users have a tendency to prefer, based on the evaluation of user and object relationship. In this research, applications in e-commerce sector and the diversity created by them have been analyzed and recommender systems are tried to be compared based on users' purchasing behaviors. In the study, the methods that are used for recommender systems are investigated and information regarding the development of these kinds of systems have been presented. In the light of many studies about different user – product relations and model establishment techniques, an application that acknowledges users' purchasing behaviors as direct product reviews of users, has been developed. In this regard, by selecting some of the recommender system methods from the literature, developed models have been compared. The model performances such as Jaccard and Cosine similarity measurement used with product-based neighborhood models, SVD and NMF methods from matrix separation methods and the Fp-Growth and PSO methods included in the association rules were evaluated by comparing within each other and also with each other. While comparing the results obtained through these results, some time periods also used in evaluation of findings as well. Score measurements have been calculated by associating re-purchases of users in the scope of all data sets per 45, 90, 365 days periods in line with determined reference day and the findings obtained via these methods. After that, in order to test the consistency of the results, the models were run repeatedly for 10 day periods and crosscheck is sustained. According to the results obtained, both in short term and in long term, in the binary data type models belonging to users' purchase behaviors, association rules of Fp-Growth and PSO obtained more successful results. Although the association rules have a better outcome relative to other methods, the time interval criterion has disclosed that their performance in itself is higher in short time intervals.