The Empirical Comparison of the Supervised Classifiers Performances in Implementing a Recommender System using Various Computational Platforms

Abstract

Recommender Systems (RS) help users in making appropriate decisions. In the area of RS research, many researchers focused on improving the performances of the existing methods, but most of them have not considered the potential of their employed methods in reaching the ultimate solution. In our view, the Machine Learning supervised approach as one of the existing techniques to create an RS can reach higher degrees of success in this field. Thus, we implemented a Collaborative Filtering recommender system using various Machine Learning supervised classifiers to study their performances. These classifiers implemented not only on a traditional platform but also on the Apache Spark platforms. The Caret package is used to implement the algorithms in the classical computational platform, and the H2O and Sparklyr are used to run the algorithms on the Spark Machine. Accordingly, we compared the performance of our algorithms with each other and with other algorithms from recent literature. Our experiments indicate the Caret-based algorithms are significantly slower than the Sparklyr and H2O based algorithms. Also, in the Spark platform, the runtime of the Sparklyr-based algorithm decreases with increasing the cluster size. However, the H2O-based algorithms run slower with increasing the cluster size. Moreover, the comparison of the results of our implemented algorithms with each other and with other algorithms from recent literature shows the Bayesian network is the fastest classifier between our implemented classifiers, and the Gradient Boost Model is the most accurate algorithm in our research. Therefore, the supervised approach is better than the other methods to create a collaborative filtering recommender system.

The Role of Machine Learning in Astronomy

“…In supervised learning [24][25][26][27], authors have an idea of what our result will look like, and therefore our goal is to train the data set in the best possible way to get the desired output. After this training is complete, authors provide the machine with new data and then this new data set is processed with the supervised learning algorithm and in the end, authors are provided with new labelled data.…”