Scalable Implementation of E Als Based Implicit And Explicit Feedback For Recommendations

Abstract

Recommender systems are now a popular tool used to suggest products, newlineservices and information to potential consumers, based on their profile of newlinepast transactions and feedback from other users that share similar interests. newlineWith the tremendous growth of users, products and information made newlineavailable on the web and the rapid introduction of new e-business services, newlineperforming many recommendations per second for millions of users has newlinebecome a necessity and a challenge. Many recommender systems suggest newlineitems to users employing collaborative filtering techniques, which process newlinehistorical records of items that the users have viewed, purchased, or rated. newlineTwo major problems that most collaborative filtering approaches have to newlineresolve are scalability and sparseness of the userand#8223;s profile matrix, which newlinehave been successfully overcome with the use of latent factor models newlinetechnique. newlineThe most successful realizations of latent factor models are based newlineon matrix factorization. Among the matrix factorization algorithms, EAlternating newlineleast squares (E-ALS) stands out because its computations are newlineeasily parallelizable. In this work we propose a methodology for comparing newlinethe performance of two parallel implementations of the E-ALS algorithm, one newlineexecuted with MapReduce in Apache Hadoop framework and another newlineexecuted in Apache Spark framework, which makes careful use of blocking newlineto reduce JVM garbage collection overhead and to utilize higher-level linear newlinealgebra operations. To evaluate their performance, RMSE is applied as the newlinemain criteria. We perform experiments to evaluate the accuracy of generated newlinerecommendations and the execution time of both algorithms, using publicly available datasets with different sizes and from different recommendation newlinedomains. Experimental results show that running the recommendation newlinealgorithm on Spark framework is in fact more efficient, once it provides inmemory newlineprocessing, in contrast to Hadoopand#8223;s two-stage disk-based newline