## Randomized coordinate descent methods for big data optimization

##### View/Open

##### Date

01/07/2014##### Author

Takac, Martin

##### Metadata

##### Abstract

This thesis consists of 5 chapters. We develop new serial (Chapter 2), parallel (Chapter
3), distributed (Chapter 4) and primal-dual (Chapter 5) stochastic (randomized) coordinate
descent methods, analyze their complexity and conduct numerical experiments on synthetic
and real data of huge sizes (GBs/TBs of data, millions/billions of variables).
In Chapter 2 we develop a randomized coordinate descent method for minimizing the sum
of a smooth and a simple nonsmooth separable convex function and prove that it obtains
an ε-accurate solution with probability at least 1 - p in at most O((n/ε) log(1/p)) iterations,
where n is the number of blocks. This extends recent results of Nesterov [43], which cover
the smooth case, to composite minimization, while at the same time improving the complexity
by the factor of 4 and removing ε from the logarithmic term. More importantly, in contrast
with the aforementioned work in which the author achieves the results by applying the method
to a regularized version of the objective function with an unknown scaling factor, we show
that this is not necessary, thus achieving first true iteration complexity bounds. For strongly
convex functions the method converges linearly. In the smooth case we also allow for arbitrary
probability vectors and non-Euclidean norms. Our analysis is also much simpler.
In Chapter 3 we show that the randomized coordinate descent method developed in Chapter
2 can be accelerated by parallelization. The speedup, as compared to the serial method, and
referring to the number of iterations needed to approximately solve the problem with high
probability, is equal to the product of the number of processors and a natural and easily
computable measure of separability of the smooth component of the objective function. In the
worst case, when no degree of separability is present, there is no speedup; in the best case, when
the problem is separable, the speedup is equal to the number of processors. Our analysis also
works in the mode when the number of coordinates being updated at each iteration is random,
which allows for modeling situations with variable (busy or unreliable) number of processors.
We demonstrate numerically that the algorithm is able to solve huge-scale l1-regularized least
squares problems with a billion variables.
In Chapter 4 we extended coordinate descent into a distributed environment. We initially
partition the coordinates (features or examples, based on the problem formulation) and assign
each partition to a different node of a cluster. At every iteration, each node picks a random
subset of the coordinates from those it owns, independently from the other computers, and in
parallel computes and applies updates to the selected coordinates based on a simple closed-form
formula. We give bounds on the number of iterations sufficient to approximately solve the
problem with high probability, and show how it depends on the data and on the partitioning.
We perform numerical experiments with a LASSO instance described by a 3TB matrix.
Finally, in Chapter 5, we address the issue of using mini-batches in stochastic optimization
of Support Vector Machines (SVMs). We show that the same quantity, the spectral norm of
the data, controls the parallelization speedup obtained for both primal stochastic subgradient
descent (SGD) and stochastic dual coordinate ascent (SCDA) methods and use it to derive novel
variants of mini-batched (parallel) SDCA. Our guarantees for both methods are expressed in
terms of the original nonsmooth primal problem based on the hinge-loss.
Our results in Chapters 2 and 3 are cast for blocks (groups of coordinates) instead of
coordinates, and hence the methods are better described as block coordinate descent methods.
While the results in Chapters 4 and 5 are not formulated for blocks, they can be extended to
this setting.