Large-scale learning with AdaGrad on Spark

Stochastic Gradient Descent (SGD) is a simple yet very efficient online learning algorithm for optimizing convex (and often non-convex) functions and one of the most popular stochastic optimization methods in machine learning today. One drawback of SGD is that it is sensitive to the learning rate hyper-parameter. The Adaptive Sub-gradient Descent, AdaGrad, dynamically incorporates knowledge of the geometry of the data observed in earlier iterations to calculate a different learning rate for every feature. In this work, we implement a distributed version of AdaGrad for large-scale machine learning tasks using Apache Spark. Apache Spark is a fast cluster computing engine that provides similar scalability and fault tolerance properties to MapReduce, but in contrast to Hadoop's two-stage disk-based MapReduce paradigm, Spark's multi-stage in-memory primitives allow user programs to load data into a cluster's memory and query it repeatedly, which makes it ideal for building scalable machine learning applications. We empirically evaluate our implementation on large-scale real-world problems in the machine learning canonical tasks of classification and regression. Comparing our implementation of AdaGrad with the SGD scheduler currently available in Spark's Machine Learning Library (MLlib), we experimentally show that AdaGrad saves time by avoiding manually setting a learning-rate hyperparameter, converges fast and can even achieve better generalization errors.