With the advent of mass-available Internet, twenty-first century observed a steady growth in web based commercial services and technology companies. Most of them are based on web applications that receive huge amount of user traffics, and generate massive amount of web usage data containing user-item interactions. We attempt to build a recommendation algorithm based on such web usage data. It is essential that recommendation algorithms for such applications are highly scalable in nature. Existing algorithms such as matrix factorization run several iterations through the dataset, and therefore may not be suitable for large web-scale datasets. Here we propose a highly scalable recommendation algorithm based on recently proposed Method of Moments (also known as Spectral Method). Our method takes only two to three passes through the entire dataset to extract the model parameters during the training phase. We demonstrate the competitive performance of our algorithm in comparison with the existing algorithms on various publicly available datasets through several empirical measures.
View on arXiv