Merlin: A GPU Accelerated Recommendation Framework

The scale of recommender system datasets in industry has grown to the point where special thought must be taken in both the preparation of the data and in the training methods used in order to avoid performance issues that can slow down the total training iteration ∗Corresponding Author Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from permissions@acm.org. IRS ’20, August 22–27, 2020, San Diego, CA © 2020 Association for Computing Machinery. ACM ISBN 978-1-4503-6201-8/20/08. . . $15.00 https://doi.org/10.1145/3292500.3330823 time by orders of magnitude. Extract Transform Load (ETL) and data preparation can take more time than training, leading to the adage that data scientists spend >75% of their time preparing data for modelling. These large datasets are required in order to enable deep learning (DL) based recommender systems which outperform traditional methods in industry settings where small differences in model performance can have a significant impact on the bottom line. Similarly, training DL recommenders whose embeddings scale beyond a single GPU requires significant expertise to implement efficiently. In this paper we present Merlin, an open source Graphics Processing Unit (GPU) accelerated recommendation framework that scales to datasets and user/item combinations of arbitrary size. The framework provides fast feature engineering and preprocessing for operators common to recommendation datasets and high training throughput of several canonical deep learning based recommender IRS ’20, August 22–27, 2020, San Diego, CA Oldridge, Lee, et al. models including Wide and Deep[3], Deep Cross Networks[14], DeepFM[6], and DLRM[8] to enable fast experimentation and production retraining. For production deployment Merlin also provides low latency, high-throughput inference. These components combine to provide an end to end framework for training and deploying deep learning recommender system models on the GPU that is both easy to use and highly performant. The Merlin framework is freely available and open source[10, 11].